All Journal Club members can pull one episode from the archive per day. (This is in addition to the new episode that will arrive every morning). If you find something on this list that interests you, email us to have it pulled and sent to you. To get the PDF of any paper, just click on the image instead.
Release Date: April 2nd 2025
In order for a diagnostic method to stand out, we need a method that has a standard unit of measure that is specific and reliable. Flow Mediated Skin Fluorescence (FMSF) can potentially fill this gap. While it has certain disadvantages, new research is finding ways to work around them. Previously, one of the main disadvantages was that FMSF failed to differentiate between vascular and mitochondrial dysfunction. But, in this paper, the authors managed to distinguish one from the other. Let’s dive in to see how this was done. So what is FMSF? It is a non-invasive test that evaluates microcirculation by detecting changes in skin fluorescence. Fluorescence is a phenomenon that occurs when light is absorbed by a molecule and then re-emitted as a different color. This happens because the light absorption excites the electrons in the molecule which causes it to glow. Nicotinamide Adenine Dinucleotide (NAD) is a coenzyme that exists in two...
Request this Episode
Release Date: April 1st 2025
Risk-sensitive foraging is a behavioral response to a “landscape of fear” that is created when a competing predator is in an area, like coyotes and wolves for example. In this case the “competing predators” are livestock guardian dogs. The weaker predator must always be aware of their surroundings for fear of being attacked and even killed, so they will usually reduce how much they forage (which means looking for food, scavenging, and hunting). Basically the predator asks itself “is attempting to eat a juicy sheep worth the risk of being killed by 3 giant sheepdogs? Probably not.” We can measure risk-sensitive foraging with a thing called giving up density (GUD). This is how willing an animal is to stay in one place in order to get more food. The idea is this: if they are in a high risk area, they will forage for less time. Alternatively, if there is lower risk, they will spend as long as they can getting more food....
Request this Episode
Release Date: March 31st 2025
In the XR world, in order to augment reality, most head-mounted devices use what’s called a Near-Eye Display (NED) system. For AR devices, this NED is paired with an optical combiner (OC). The job of an OC is to reflect computer-generated images into your eye while allowing real-world light to pass through. Prism optics, waveguides, and Freeform lens systems are the usual suspects here, and each has its pros and cons. Prisms are light and cheap, but they give you a narrow field of view (FOV). Waveguides are sleek but struggle with chromatic aberration and light loss. And freeform optics offer great visuals but tend to be bulky and thick. Across the board, these systems are fundamentally static. They lack real-time adaptability. Want to switch from full immersion to light overlay or no overlay at all? Well, you're out of luck. So what if the optical layer itself could be made to change? What if you could “tune” it in real...
Request this Episode
Release Date: March 30th 2025
In 2020, a bank manager in Hong Kong received a call. It was from the senior director of a company that regularly did business with the bank. The person explained that the company was about to make a major acquisition, and he requested $35 million to be transferred to the seller's account. The money was sent, and everything seemed normal. Not long after, the bank employee was shocked to learn that he had not spoken to the director at all. He had fallen victim to a Deepfake scam. Today we’ll look at how researchers are trying to stay ahead of the growing threat of Deepfake audio using a new AI detection model called Sonic Sleuth. We’ll look at how they created the model, what feature extraction tools were most effective, and if Sonic Sleuth might help us determine the difference between real and fake. Before we look at the model, let’s talk about what makes Deepfake audio so dangerous and hard to detect....
Request this Episode
Release Date: March 29th 2025
Consider training a language model on a corpus where shorter sentences are overrepresented. If the model updates its parameters using mini-batches drawn randomly from the full dataset, shorter and syntactically simpler examples will dominate the early gradient estimates. These examples might favor certain token predictions or structural patterns that are not representative of the broader language distribution. As a result, the model parameters will be nudged in directions that overfit these simpler forms. This doesn’t just introduce noise, it creates a directional bias in the gradient field, pulling the optimization process toward a region that minimizes loss on an unbalanced subset of the data. Even as training continues and more varied examples appear, the model may be stuck in a basin shaped by these early biases. This phenomenon, while hard to detect in raw training curves, has been observed in practice in domains like...
Request this Episode
Release Date: March 28th 2025
Draw 6 dots in the shape of a hexagon, and connect the dots with a line. Now draw the same hexagon next to that hexagon, so that they tessellate (fit together with no gaps between them). Keep growing your hexagons up down and side to side until it resembles a honeycomb lattice, or chicken wire. Congratulations! You’ve drawn graphene. Each dot represents a carbon atom and each line is the covalent bond between them. How do you even make a material that's 1 atom layer thick? The answer is surprisingly simple: scotch tape. Professor Sir Konstantin Novoselov (Kostya) was a PhD student working under Professor Sir Andre Geim at the University of Manchester in 2004 when they made a surprising discovery. Geim encouraged the lab group to spend Friday afternoons playing around with different research ideas. One such idea involved trying to isolate a material that was just 1 atom layer thick. Here’s how they did it: They obtained a lump of...
Request this Episode
Release Date: March 27th 2025
Esophageal cancer is extremely hard to treat due in large part to one particularly diabolical factor, tumor heterogeneity. As you might imagine, tumors are not all the same. With a wide variety of cancer types, it would stand to reason that there would be variety in tumors. That would be an understatement. ESCC tumors can vary wildly from one another. Take for example a patient with esophageal cancer with two malignant tumors. These tumors may have variations in cell compositions, DNA, and protein expression. Even if both tumors were caused by the same genetic mutation in the host DNA, they may develop completely different cell structures, express completely different biomolecules and even have different types of additional genetic mutations. The tumors may also react to treatments differently, putting stress on patients and increasing the potential for the cancer to return. On top of these variations, recent studies have...
Request this Episode
Release Date: March 26th 2025
The researchers performed a channel capacity analysis, a test to determine the theoretical maximum rate at which information can be transmitted, measured in bits per second. To calculate the overall channel capacity for THz transceivers in paint, the total bandwidth is divided into narrow sub-bands, and their individual capacities are aggregated. This is possible due to the high frequency selectivity at THz frequencies; in other words, the system is capable of responding to specific frequencies while ignoring others. The division into sub-bands is also helpful to mitigate potential interference from non-white molecular noise, which varies across the frequency spectrum. In the simulation, they used a 2 mm thick layer of “Titanium white” paint; this is thicker than most paint layers and was used to represent a worst-case scenario. They then modeled three different transceiver burial depths, 0.5 mm, 1 mm, and 1.95 mm and checked...
Request this Episode
Release Date: March 25th 2025
In classic game theory, players make rational decisions based on perfect information. But in real life, especially in complex environmental crises, information is limited, and rationality is bounded. That's where evolutionary game theory excels. It acknowledges that players don't instantly arrive at optimal strategies but rather adapt their approaches over time based on what they learn. The researchers created a model that gives each player two options: Japan can choose to either discharge the wastewater, or not. Other countries can choose to sanction Japan, or not. The JFA can choose to support the Japanese government, or not. Each choice carries costs and benefits. For Japan, discharging wastewater reduces storage costs but could damage their international image and lead to potential economic penalties. Other countries might gain political points by sanctioning Japan but doing so could disrupt valuable trade relationships. The JFA could...
Request this Episode
Release Date: March 24th 2025
For a long time, pure reinforcement training has been the go-to method. It works by rewarding correct actions and penalizing incorrect ones, encouraging the model to maximize positive outcomes over time. More recently, Reinforcement Learning from Human Feedback has helped improve training time and alignment with human preferences, predictions, and judgment. Chatbots like ChatGPT, Claude, and Gemini have used this method to refine their conversational abilities over time. Early chatbot responses were factual but often rude or confusing. In the world of autonomous vehicles, self-driving cars also require a level of knowledge that allows them to overcome small unpredictable obstacles that were never incorporated into the training dataset. By incorporating human feedback, AI models could become more natural and intuitive. In most RLHF systems, you’d train your model first and then provide feedback to it after it’s already up and running. The authors did something different: they built a system that could provide a form of “pre-feedback” to the model well before training had taken place. So rather than...
Request this Episode
Release Date: March 23rd 2025
An in-person ultrasound involves a patient lying in various positions to optimally target the organs or body region to be screened. The sonographer holds a transducer, a device that looks like a cross between a universal TV remote and a grocery store check out scanner. The transducer makes contact with the skin, and sends the soundwaves into the body, where they bounce off soft tissues. The echo bounces back and is received by the transducer, which sends that information back to a computer to generate the images. A gel is applied to the region of the body that the scanner will touch to allow smooth movement of the transducer along the skin, and to eliminate any air pockets (between the skin and the device) that might interfere with the sound waves. Some scans can be as short as five minutes, while others can last thirty minutes. That's a lot of time to be in intimate proximity with anyone, let alone someone...
Request this Episode
Release Date: March 22nd 2025
In the last few years, LiDAR and Structure from Motion photogrammetry (SfM) have made progress in terms of quantifying biomass in dryland ecosystems. Unfortunately, we still need things at a finer scale, especially when it comes to grasslands. In grasslands, plants are thin and tall, grow densely packed, and have a lot of built up litter. These factors make it necessary to have high-resolution tools that allow us to create an accurate 3D rendering of the plants in an area. The existing systems work by taking images and generating a kind of 3D landscape called a point cloud. The data in the point cloud can then help us calculate the parameters necessary to estimate AGB. But, how accurate are they? Before we ask that question, let’s explain how these remote sensing systems work. LiDAR is considered active remote sensing. It works by emitting laser pulses and measuring the time it takes for them to return. This can create precise...
Request this Episode
Release Date: March 21st 2025
An NIC or Network Interface Card, does exactly what its name suggests. It manages a device's network connection, letting it connect, share resources, and communicate either locally or remotely. A smartNIC goes a step further, it does all of that and more. They are designed to handle high-traffic operations and they come with features like packet processing, load management, and programmable processing units. These units are capable of running custom applications. Traditionally, some networking tasks placed a heavy burden on a server’s processing power. With SmartNIC’s acting as intelligent offload engines, these functions can now be handled directly at the network interface, freeing up computational power for other critical operations. So why is this important? Well, most solutions to the cold start problem have been tackled at the server level, by throwing more CPU and GPU at the problem. But, even with...
Request this Episode
Release Date: March 20th 2025
In 1952, Bell Labs first developed Audrey, a system that could transcribe spoken numbers. In the 1980s, the development of the Hidden Markov Model created systems that were capable of recognizing thousands of words. This evolved into what is called the “hybrid approach,” in which three models, a lexicon model, an acoustic model, and a language model were forced into alignment to generate a more accurate transcription. With the advent of deep learning, we're able to directly map input audio to text with greater speed and higher accuracy. However, when it comes to recordings, getting an accurate transcription is only half of the picture. Recorded speech contains a variety of non-textual information, such as tone, loudness, or even accents. The field of audio processing has been evolving in parallel to these transcription tools. In today's paper, the authors took a look at a variety of modern tools for analyzing speech. They placed the...
Request this Episode
Release Date: March 19th 2025
Last summer, as the Olympic games came to Paris, all eyes were on the Seine. It was filthy. Years of untreated wastewater overflowing into the river had left it full of so much bacteria and so many viruses that it was genuinely dangerous for the athletes to jump in. And sewage wasn’t the only problem. The river was also full of antibiotics. Over a dozen different types were found in the waters. This isn’t just a problem in France. It’s far more widespread. Every major waterway suffers from this kind of contamination. But what are we to do about it? Treatment plants aren’t doing their job. Chemicals aren’t going to help. So what other options are on the table? Ideally we need teeny tiny machines that can swim around, cleaning up rogue antibiotics, and keeping rivers clear. Microrobots! In this paper, that’s exactly what these researchers came up with! In this episode of journal club, we’ll be looking at how they designed and built...
Watch the Trailer
Release Date: March 18th 2025
This story takes place at what is called an XFEL facility. In an XFEL (which stands for X-ray Free Electron Laser) lab, researchers regularly fire ultrafast, high-intensity X-ray pulses at microscopic crystals to determine their atomic structure. A core part of that process is, as I mentioned, Bragg Peak detection. A Bragg Peak is a bright spot in a diffraction pattern. It’s formed when X-rays scatter off a crystal’s atomic lattice and interfere constructively at specific angles, in accordance with something called Bragg’s Law (thus the name). These peaks hold the key to determining a crystal’s atomic structure because their positions and intensities reveal how X-rays interact with the crystal at the atomic level. A modern XFEL can produce up to a million images per second, and they all need Bragg Peak detection run on them. But there's the problem: it can't handle that kind of speed and scale. Why? Well, existing...
Request this Episode
Release Date: March 17th 2025
Memory leaks occur when dynamically allocated memory is never properly released. This leads to gradual accumulation of inaccessible memory that cannot be reclaimed by the operating system. This is particularly problematic in C and C++, where memory management is manual, meaning there’s typically no automatic garbage collection in most implementations of those languages. So developers must explicitly allocate memory using functions like “malloc” or “new” and then free it with “free” or “delete”. When memory is allocated but never freed, it remains reserved for the duration of the program's execution, even if it is no longer reachable. Over time, this results in progressively increasing memory usage. If you’re watching the memory utilization on a computer executing this kind of program you can literally see it creeping up little by little over time. Eventually this leads to performance degradation and, in the worst cases...
Request this Episode
Release Date: March 16th 2025
We are running low on solutions to combat superbugs. Rather than sticking with antibiotics, we’re starting to turn our gaze towards gene therapies. To this end, the authors of this paper have reviewed the latest applications of CRISPR: Clustered Regularly Interspaced Short Palindromic Repeats. They’re looking specifically at how it can be used to modify the genes of superbugs and make them more susceptible to antibiotics. Today on Journal Club, we will be talking about antimicrobial resistance (AMR), what CRISPR is, and why CRISPR could be the ultimate weapon in this battle. In 2016, a 70-year-old from Nevada died in the hospital, overcome by an infection from the bacteria Klebsiella pneumoniae. Why is this significant? The specific strain of Klebsiella had evolved into a superbug. It had developed resistance to all the antibiotics available in the United States. This means that no known medication was capable of...
Request this Episode
Release Date: March 15th 2025
Aerosol optical depth or (AOD), is the metric for measuring the density of aerosols in the atmosphere. The AOD is determined by measuring the amount of light scattered or absorbed by aerosols; in other words, how much direct sunlight is prevented from reaching the ground. Lower values below 0.05 indicate a clear sky, whereas an AOD above 2 or 3 indicates a high particle concentration. The AOD is measured globally through an array of satellites equipped with radiometers (sensors that measure the intensity of radiant energy). By comparing how the AOD changes at various wavelengths of light, these satellites can also determine the Ångström Exponent, a measurement of particle size. When the effects of aerosol are substantially different for red and blue wavelengths, the resulting Ångström Exponent is higher. Lower exponent values typically indicate natural aerosols such as sea spray or mineral dust, whereas higher values often...
Request this Episode
Release Date: March 14th 2025
Similar Integer Coding incorporates an awareness of service similarity. Instead of treating each service as an entirely distinct option, this approach groups services with comparable QoS attributes into clusters before assigning integer codes. These clusters are formed based on a similarity measure, specifically cosine similarity, which evaluates how close two services are in terms of key performance metrics. Cosine similarity is a mathematical technique that measures the angle between two vectors in a multi-dimensional space, rather than their absolute distance. In the context of service composition, each service can be represented as a vector of its QoS attributes, where each attribute (execution time, cost, reliability) corresponds to a dimension. The cosine similarity score between two services is computed as the dot product of their vectors divided by the product of their magnitudes, resulting in a value between -1 and 1. A similarity...
Request this Episode
Release Date: March 13th 2025
To understand how it works, and what benefit it has to our environment, we need to look at the basics. Two electrodes: an anode and a cathode, are placed into a water source. Let’s use a jar as an example. Each electrode is connected to a power source, with the anode attached to the positive terminal and the cathode to the negative. When power is supplied, a reaction occurs, splitting the water into its two elements: hydrogen and oxygen. The hydrogen is captured, while the oxygen is either stored or released into the atmosphere. This entire process is known as green hydrogen production, and when applied to wastewater, it provides a viable way to create a clean fuel source from water that would otherwise be discarded. But there are two major challenges with this process. First, chemical reactions that can occur in wastewater can produce hazardous gases beyond just hydrogen and oxygen. Second, to test these processes effectively...
Request this Episode
Release Date: March 12th 2025
The year was 1986, and the Department of Defense had a problem. They were interviewing Software Development agencies in order to award contracts, and they needed a way to pick a winner. All the dev-shops interviewed quite well. They all had expansive portfolios. They’d all write very convincing proposals, they’d give impressive presentations, and they all seemed to know what they were doing. The issue was, the Department had been fooled in the past. Previous contractors had overstated their abilities. And many, it turned out, weren't actually capable of even a fraction of their claims. Given a contract, they’d try! They’d certainly try!...to build what was asked for, but they wouldn’t succeed. Instead, they’d miss deadlines, or deliver buggy applications, or just give up and abandon the project. So, the Feds were left with an interesting question: How do you evaluate a Software Engineering team that you just met? How do you...
Request this Episode
Release Date: March 11th 2025
So how can we allow a surgeon to control nano-sized objects with precision and intuition, without the huge costs and steep learning curve? The authors of this study aimed to find out just that. A solution that incorporated a wearable device, an electromagnetic generator, and an AI planner, enabling 3 dimensional control of a nanorobot. This device needed to consistently extract physiological signals from the user while accurately maneuvering the robots in real time. The researchers proposed a Fully Integrated Wearable Control System (FIWCS) that detects surface electromyogram signals (sEMG) and gyroscopic rotations through gestures like waving and wrist rotations. They also incorporated a microphone to allow surgeons to initiate tasks using voice commands. Tests were then run using an optical microscope to monitor the speed, rotation and task execution. To understand the true scale of this innovation, we need to see how these nanorobots...
Request this Episode
Release Date: March 10th 2025
Side-channel attacks exploit unintended information-leakage from a system’s physical operations. These leaks can emerge in different forms, such as timing variations, electromagnetic emissions, or even acoustic signals. One of the most effective and well-documented forms of side-channel attacks is power analysis, which observes fluctuations in a device’s power consumption as it processes cryptographic operations. Every computational step, from simple arithmetic to complex encryption functions, requires electrical energy, and the precise way that energy is used can reveal patterns about the underlying operations. Power analysis attacks take advantage of these patterns to infer secret information, such as cryptographic keys, without needing to directly access the encrypted data. Traditional power analysis techniques include simple power analysis (SPA) and differential power analysis (DPA)...
Request this Episode
Release Date: March 9th 2025
In this episode, we’ll explore the innovative way that the authors harnessed the FAPI Nanoparticles to positively impact theranostics in cancer treatment. We’ll walk through what the FAPI Nanoparticle is, what makes it so useful, and how the researchers tested it. We’ll then discuss how it may be able to revolutionize the treatment of tumors. So let’s start this story by meeting another molecule that sits at the center of all of this. Fibroblasts activating protein, or FAP. Not FAPI, that’s something else. More on that later. FAP is a protein expressed on the outer membrane of close to 90% of the tumors growing on organs. These molecules have very specific molecular structures that are used to locate tumors in living tissue. In essence imaging technologies detect tumors in the body by looking for specific biomarkers like FAP that tumors express on their outer surfaces. Like sprinkles on a cake. FAP is such...
Request this Episode
Release Date: March 8th 2025
Data leakage occurs during training, when information that should be unavailable to the model inadvertently influences its learning process. Think of it as contamination. Your testing/evaluation data somehow got into the training data. As you can guess, this can lead to overly optimistic performance metrics that do not translate into real-world effectiveness. This issue is particularly insidious because it masquerades as a highly successful model, with impressive training and validation results that collapse once the model is deployed. The problem is not that the model has failed to learn meaningful patterns, but rather that it has learned patterns it should never have had access to in the first place. As a result, the model does not actually generalize but instead exploits unintended shortcuts in the data. So how does this happen? There are several ways, and they stem from flaws in your data pipeline...
Request this Episode
Release Date: March 7th 2025
In the fast-paced world of Formula One racing, engineers work tirelessly to construct a vehicle that can achieve the impossible, breakneck speeds with razor-sharp control and precision. Unfortunately, they’re being driven by humans. And humans make mistakes, so crashes are inevitable. To mitigate serious injuries, racetracks are lined with barriers meant to soften the impact of the crash. But how do we know if those barriers are good enough to save a driver’s life? Well, we test them, over and over again. Today, we’ll examine the testing procedures used on tire barriers at Formula 1 race tracks. We’ll start by looking at standard Finite Element (FE) simulations, then we’ll explore a novel Soft-Body physics simulation, and finally compare both to physical crash tests. Before diving in, we need to consider the scale of vehicle testing, especially in high-speed racing. We’ve all seen those advertisements where a...
Watch the Trailer
Release Date: March 6th 2025
Let’s say you have now collected some mosquitoes and preserved them. What’s next? Once DNA is collected from a substrate, in this case the blood that a mosquito had for lunch, that DNA must be extracted, amplified, sequenced, and compared against genomes that we already know. A popular method for DNA amplification is PCR. In PCR (Polymerase chain reaction): Extracted DNA is heated so that it separates into two strands. Short synthetic DNA fragments… called… primers are added. Using the primers as templates, desired DNA is amplified through multiple rounds of DNA synthesis. Once there is enough of the DNA it can be sequenced and we can figure out what it belongs to. So, let’s get into what the authors did and how it turned out. The authors set up light traps to collect mosquitoes and see what level of vertebrate diversity they could survey in just 4 nights. Mosquitos were...
Request this Episode
Release Date: March 5th 2025
The paper introduces a workload classification system called TRM-MAAN. It is designed to assess pilot workload in real time using physiological data. The system relies on two primary signals: EEG: which measures electrical activity in the brain. EMG: which captures muscle activity. Both of these signals offer insight into a pilot’s cognitive and physical state but integrating them into a single, reliable model is challenging. EEG and EMG operate on different time scales and frequency ranges, meaning that traditional fusion techniques often fail to align them properly. Moreover, workload is inherently dynamic, changing continuously throughout a flight, so any meaningful classification model must account for temporal dependencies rather than treating data points in isolation. To address these issues, the framework uses a Transformer. The reason for choosing a Transformer (over other deep learning architectures) lies in...
Request this Episode
Release Date: March 4th 2025
Real-world farming isn’t as neat and simple as a library of images. Nature is unpredictable. No amount of training data will account for variables like lighting, position, and bunch size. These variations make it much harder for AI to identify the tomatoes accurately 100% of the time. And when you factor in the complexity of the plant’s structure into the mix, things get even messier. This is where Multi-Object Tracking (MOT) technology comes in. It’s a set of algorithms designed to track multiple objects while they, or the camera, are in motion. In MOT systems this is known as re-identification. Think of the technology that allows commentators to track each player on a football field during a game. While the human eye perceives this as smooth tracking, it’s actually a series of frames fed into an algorithm, along with their corresponding location within three-dimensional space. This 3D space is what’s known...
Request this Episode
Release Date: March 3rd 2025
ASO is rooted in molecular dynamics, a branch of physics that describes how atoms interact under forces like attraction and repulsion. There are principles that govern the movement and behavior of atomic particles, and they form the basis for ASO’s search strategy. Each candidate solution is represented as an atom within a simulated environment. The algorithm then mimics two fundamental forces: Interaction forces: These describe the attraction and repulsion between atoms. Interaction forces are based on the Lennard-Jones potential, a well-known function in physics that describes how atoms attract each other at long ranges and repel each other at short distances. In this context, this means that candidate solutions close to one another in the search space will experience repulsion, preventing overcrowding and encouraging diversity, while those farther apart will experience attraction, helping the algorithm converge toward...
Request this Episode
Release Date: March 2nd 2025
Every parents’ worst nightmare: Your child has just been diagnosed with a rare condition, and you’re finally home from the hospital. Time to give them their medication. The pharmacy only had adult tablets, so you’ll need to break them in half. The two halves are not equal, and you notice some collateral damage on the kitchen table. Given the option between underdose and overdose, you choose option 3: Make some coffee. As you stand in front of the machine dialing-in the amount of caffeine the situation calls for, you think to yourself: Why isn’t there a machine that can dial-in the right amount of medication for my child? …well...maybe there is. In today’s paper, the authors are assessing the feasibility of using 3D printers to modify adult tablets into smaller, patient-specific doses. If they’re successful, they’ll have proved that it’s possible to downsize a tablet to a...
Request this Episode
Release Date: March 1st 2025
Traditional logic operates in absolutes, where something is either true or false, present or absent, good or bad. But many real-world problems don’t fit neatly into binary classifications. Fuzzy logic provides a way to represent uncertainty by allowing variables to exist in degrees rather than absolutes. Instead of saying a coffee brand is either "high quality" or "low quality," fuzzy logic assigns a membership value to quality, indicating the extent to which a brand belongs to that category. This approach more closely mirrors human reasoning, where preferences and perceptions are rarely black and white. Multi-criteria decision-making, or MCDM, is a framework designed to evaluate and rank alternatives when multiple, often conflicting, factors must be considered simultaneously. Unlike single-factor decision-making methods that prioritize one variable, such as cost or performance, MCDM recognizes that real-world decisions...
Request this Episode
Release Date: February 28th 2025
On October 4th, 1957, the Soviet Union launched Sputnik: the first satellite ever put into orbit. This, to put it bluntly, scared the crap out of the Americans and propelled the U.S. into the space race. In the days and weeks after launch, the U.S. became obsessed with tracking Sputnik's movements. Where is it now? Ok, how about now? And now? In order to track it, they used Doppler shift measurements. Radio receivers on the ground would listen to Sputnik’s beeping signal, and analyze how its frequency changed as it moved. The resulting measurements allowed engineers to calculate its exact orbit and predict its future positions. The U.S. got quite good at this, and this led to a realization. If we can track a satellite in space with ground-based measurements, then the converse is also true. If we had multiple satellites in space at one time, then we could use them to track objects on Earth. Within four months...
Request this Episode
Release Date: February 27th 2025
Herbal medicines often have many bioactive compounds, working simultaneously on many different parts of the body, making it difficult for researchers to identify the actions of specific molecules. Mass Spectrometry Imaging (MSI) techniques have provided a visual window into the molecular realm of herbal medicines, enabling researchers to identify locations of useful compounds and observe the chemical mechanisms of pathology. Molecular level imagining can greatly improve the quality control of the herbal medicine landscape by helping to standardize cultivation and processing methods, ensuring safe and reliable medicines for all. In this episode of Journal Club, we’ll explore the three major methods of MSI, present how researchers use MSI techniques in herbal medicine analysis, and see how they can be applied to quality control improvement. But first, let’s start things off by looking at how herbal medicines work....
Request this Episode
Release Date: February 26th 2025
Delaunay triangulation constructs a set of triangles connecting the points in such a way that no point is unnecessarily enclosed within any triangle. This can be difficult to picture, so let’s take a moment to visualize it. You’ve got a 3D space, and a bunch of points floating in it. A Delaunay triangulation would try to connect these points with triangles in a way that avoids skinny or highly stretched triangles, instead favoring triangles that are as close to equilateral as possible. It doesn't add points that aren't there, and it doesn't move any of your existing points, it just figures out the way to connect the existing points such that it maximizes the minimum interior angle across all triangles. Remember that an equilateral triangle has the largest possible minimum-angle (in other words, its smallest angle is the largest it could be). As you would expect Delaunay triangulation is well-suited for...
Request this Episode
Release Date: February 25th 2025
Traditional FIFO (First-In-First-Out) are ill-suited for telephony because customers do not passively wait in line. They either hang-up and redial, or they give up entirely. This leads to something called a “retrial phenomenon”, where unserved customers retry the system in an attempt to obtain service. Sometimes they do this immediately, sometimes they do it after a delay. If this isn’t managed properly, retrial traffic can become self-perpetuating. It causes waves of congestion that overload the system, creating a kind of cyclic doom spiral that takes everything down. To make it more complicated, servers are not merely passive responders to inbound requests; they also initiate outbound calls. Many helplines proactively reach out to clients (for follow-up care, appointment reminders, or callbacks). This means that server availability is not solely determined by inbound traffic but also by internally...
Request this Episode
Release Date: February 24th 2025
Geesh...is every computer vision paper just another deep learning classifier? It can certainly seem that way. There are so many projects built on top of Yolo, ResNet, and VGG that it can be tempting to believe that that’s all there is...that there’s nothing else going on in that field. So today, we’re going to look at something different. A computer-vision problem that’s being solved in a way that we rarely see. Instead of a normal deep-learning classifier, the authors built a system that performs a deterministic traversal through a directed graph in order to arrive at a class label. Why? Because it turns out that traditional CV models are not a panacea. They have a number of drawbacks that make them less suitable for specific applications. In today’s episode we’re going to talk about how most CV models work, why they’re not ideal for certain use-cases, how this alternative method is better suited, and then how...
Request this Episode
Release Date: February 23rd 2025
We have to start with a Probability Density Function (PDF). Not to be confused with a Probability Distribution Function, which would show us a distribution (the probability of all possible values a random variable can have). A Probability Density Function shows us the likelihood of a variable having a value within a specific range of values, like rain height measurements through the summer. A higher density of rain height measurements that fall in the specific range, the more likely it is that rain will fall in that range in the future. The Density function is the one that we are interested in today. The function is the equation that data is fed into to calculate the probability spread. The visual expression of this function is a line graph with a parabolic arc. Let's use the famous bell shaped curve as an example. The middle of the curve represents the highest density of data points. The X axis is the...
Request this Episode
Release Date: February 22nd 2025
Unlike internal combustion engines, which burn fuel to produce mechanical energy, fuel cells generate electricity through an electrochemical process. They pull electrons from hydrogen atoms, creating an electric current, and then combine the hydrogen with oxygen to produce water. No flames, no emissions. There are several types of fuel cells, in two big groups: low-temperature and high-temperature. Within that latter group are several subtypes; the one we’re looking at today is called SOFC: Solid Oxide Fuel Cells. SOFCs rely on a solid ceramic electrolyte (usually a kind of zirconia called YSZ) to transport oxygen ions from the cathode to the anode. They run very hot: between 600 and 1000 degrees Celsius. This allows for greater fuel flexibility, enabling them to operate on hydrogen as well as hydrocarbon fuels like methane or even ammonia....
Request this Episode
Release Date: February 21st 2025
HPC engineers aren’t like most devs. They work in an environment that is decidedly...weird. It’s an environment where languages like Fortran abound (and are actually quite popular). An environment where batch jobs can queue for days, where computers are cooled by water instead of fans, where performance is measured in teraflops, storage is measured in exabytes, networking speeds are faster than most SSDs, and job failures can be caused by cosmic rays flipping bits. HPC is just a whole other universe. And it’s a universe that uses an almost unbelievable amount of power.High-performance computing centers are among the most power-hungry facilities in the world. A single supercomputing installation can draw upwards of 20 megawatts of juice. That’s an amount comparable to the total electricity consumption of a small city. The JUPITER system planned for Germany is expected to require an amount of power roughly...
Request this Episode
Release Date: February 20th 2025
You’ve done it. It took you months of work, but you’ve done it. You trained the perfect computer-vision model. Maybe it's a CNN, maybe it’s a vision-transformer, who cares. It’s yours, it works, it’s accurate, and you’re ready to roll it out. There’s only one problem: it’s big. Real big. Way bigger than you anticipated. Best case scenario: inference is going to be slow.Worst-case scenario: you won’t be able to deploy to the device it’s supposed to run on. So, what are your options? You’ve got quite a few: from network decomposition, to quantization, to knowledge distillation. And in today’s paper, the authors introduce another method to consider. They call it LFRNS, the Low Functional Redundancy-based Network Slimming method. And they say it’s your best way forward. The authors claim that (using this technique) they were able to cut the parameters of computer-vision models by more than half, with only a 1% loss in accuracy. How? Well...
Request this Episode
Release Date: February 19th 2025
If you watched the HBO miniseries “Chernobyl” a few years back, then this is a story you already know well. In 1986 there was an explosion in Reactor 4 at the Chernobyl Nuclear Power Plant in modern-day Ukraine. This accident released massive amounts of radiation, and created an environment that was lethal not just to humans and animals, but also to machines. You see, when Soviet engineers sent in remote-controlled robots to aid in the cleanup, many of them failed almost immediately. The radiation fried their electronics. It caused their systems to glitch, their sensors to malfunction, and their circuits to burn out. Some of the robots just stopped moving, others spun around aimlessly, and others lost communication with their operators. This led to one of the darkest moments in an already tragic event. Officials decided to send in what they termed "bio-robots" instead. Of course, these weren't robots at all...
Request this Episode
Release Date: February 18th 2025
A VAE is a deep learning model designed to learn compact representations of complex data while preserving its underlying structure. Traditional autoencoders simply compress input data into a lower-dimensional space and then attempt to reconstruct it. VAEs go further. They incorporate a probabilistic component that allows for more flexible and expressive representations. This makes them particularly well-suited for applications where data exhibits variability and uncertainty. Instead of mapping inputs to a single fixed latent representation, a VAE learns a probability distribution over possible representations, enabling it to generate new data points and make more nuanced predictions. Its architecture consists of two primary components: an encoder and a decoder, connected through a latent space. The encoder takes high-dimensional input data (such as a user’s multi-criteria ratings) and compresses it into a...
Request this Episode
Release Date: February 17th 2025
If you took a Psych, Econ, or Poli-Sci class in college, there’s a good chance you learned about Prisoner’s Dilemma. It’s a classic scenario: Two people are arrested for a crime and taken to jail. They’re held in separate cells so they can’t communicate with each other. The detectives interrogate them separately and present them both the same offer: If neither of you confess you’ll each get a one-year sentence. If one of you confesses, that person will go free and the other person will get a three-year sentence. If you both confess you’ll both get two years. What would you do? The purpose of this scenario is to highlight the conflict between self-interest and cooperation, and to show that when cooperation breaks down (or communication is unavailable) there can often be a worse outcome for all parties involved. This is true of Prisoner’s Dillemma, Stag Hunt, Tragedy of the Commons, Volunteer’s Dilemma and...
Request this Episode
Release Date: February 16th 2025
In this paper, the authors answer an important question: What can you do when you've got an optimization problem, and neither rule-based approaches, mathematical techniques nor heuristic algorithms are able to do the job? They've come up with a novel solution: a two-stage model. In the first stage, their system constructs a graph-based representation of the ADN, where power flow data and network topology are mapped into a queryable structure. This allows for easy identification of feasible load transfer paths. In the second stage, it determines the optimal sequence of switch operations, while minimizing congestion, reducing power losses, and maintaining network stability. For the first stage, they are using a database called Neo4j, and for the second stage they’re using a model called a Dueling Deep Q-Network (Dueling DQN). Let’s dive into what both of those are, and how they’re being used here...
Request this Episode
Release Date: February 15th 2025
A simple perceptron is the most basic type of artificial neural network. It consists of a single layer of neurons that directly map input features to an output. Each neuron in a perceptron receives multiple numerical inputs, applies a corresponding weight to each, sums the weighted inputs, and then passes the result through an activation function that produces an output. This is typically a step function that produces a binary output. This means a perceptron functions as a linear classifier, capable of distinguishing between two classes only if they are linearly separable (meaning they can be divided by a straight line or a hyperplane). While these are effective for simple classification problems like AND or OR logic gates, a single-layer perceptron fails to model more complex relationships where data points cannot be neatly separated by a single decision boundary. A multilayer perceptron (MLP) extends the capabilities of a simple perceptron by introducing one or more hidden layers between the input and output layers. These hidden layers contain...
Request this Episode
Release Date: February 14th 2025
We’ve all been in this situation before. We’re buying a shirt online, and we’ve found one that looks perfect. It’s the right size, it’s the right cut, and it's the right color. The only thing we’re not sure about, is its quality. How can we know if it’ll last forever, or disintegrate in the first wash? Will the photos tell us that? Or how about the reviews? Or, does price equal quality? What we could use is a rating system for clothing durability. Having that would allow us to buy more confidently, and to know our clothes will last. In today's paper, the authors present the groundwork for just such a system. One that can be applied to all types of garments. We’ll start by exploring the variables that are most important to garment durability, look at how the researchers went about testing them and then look at how they used the data...
Request this Episode
Release Date: February 13th 2025
You’ve seen this trope before, I’m sure. There’s caution-tape around a crime-scene. Detectives and investigators swarm about, looking for evidence. One of them is dusting for fingerprints, and eventually they find something! A finger print on a door-knob, or a glass, or on a window. They run the prints, and find the bad guy. Voilà, happy ending. That kind of story fits nicely into an episode of SVU, but it doesn’t actually reflect much of what the real process is like. In the real-world, there are a number of issues with fingerprints. For starters, they’re everywhere and on everything, in various states of decay. Every household appliance, and remote control, and phone, and computer, and table-top, and light-switch, and cupboard is covered in them. So the process of dusting for prints isn’t like looking for a single print or two in an otherwise clean room. The opposite: the room is full of prints, and you’re trying to figure out...
Request this Episode
Release Date: February 12th 2025
Facial recognition is so ubiquitous that we rarely stop to think about how it actually works. If I hold my phone up to my face to unlock it, what is the phone actually doing? Is this a deterministic algorithm? Is it a machine-learning model? How does it work? And in the rare case that it struggles to recognize me, why is that? In today’s paper the authors introduce a new form of facial recognition that overcomes some of the problems of the previous versions. The challenge for us is to understand the context of their contribution. We’re not going to be able to grok what these authors are doing unless we first take a look at what came before. So we’re going to start by going back. All the way back to when facial recognition started. We’ll explore what it was like in its fledgling phases. Then, we’re going to walk forward, through generation after generation of the technology to see how it has...
Request this Episode
Release Date: February 11th 2025
In simulated annealing, the algorithm starts out “hot” and slowly cools off. When the system is metaphorically "hot," it can make moves that may temporarily worsen the solution. And as the temperature decreases over time, the algorithm becomes more conservative, gradually focusing on refining the best solutions it discovered. The key innovation here is the controlled decay of exploration. This balances the need for global search early in the process, with fine-tuned exploitation as it converges. Let’s go back to the example of someone walking around a neighborhood. With Simulated Annealing, they’d jump from block to block in the beginning, randomly checking prices all over the area. And then, once they found the cheapest blocks, they’d get more conservative and focus on finding the cheapest house on one of those blocks. MOSA extends the principles of simulated annealing to problems where multiple conflicting objectives...
Request this Episode
Release Date: February 10th 2025
I find it's useful to think of "Explainability" as a continuum. On one end, you have purely deterministic, single-threaded, procedural code: Do this, then that, then this, then that. When you run the program, you'll see it do those things. And conversely: it’s always easy for you to determine why it's doing the things that it's doing. It only takes the steps that are clearly expressed in the instructions that you gave it. In the middle of the continuum are things that are probabilistic but still predictable. A function with some randomness, or even a machine learning model with well-calibrated uncertainty estimates. The system runs, and you might not know exactly what a function is going to return, but you do know that it's going to be within a range of values. On the far end of the continuum, you have deep neural networks and other high-dimensional, adaptive...
Request this Episode
Release Date: February 9th 2025
The vast majority of the papers we discuss are about pushing technology forward. A new way to solve a problem, a new trend in an industry, or a new tool that hasn’t found its use-case yet. But every once in a while, we come across a paper that is doing the opposite. A paper that’s saying “Wait, turn around, we’re headed in the wrong direction!” Today’s paper is in that latter group. The authors are making the argument that Diffusion Models are being misused. They say that diffusion models are being treated as if they’re upgraded GANs, but in reality they’re fundamentally different; and that difference makes them unsuitable for some of the use-cases to which GANs were being applied. It’s an interesting case that they’re making. And it gives us an excuse to dive into both GANs and Diffusion Models, learn how they work, how they’re the same, and how...
Request this Episode
Release Date: February 8th 2025
The title of this paper is really good at two things: 1) Intimidating the reader. 2) Itemizing all the concepts we need to understand before we’re able to understand this paper. #1 is a shame, but #2 is actually useful. Today we’re going to use this horrible title as a roadmap. A checklist of all the pieces of context we need to gather together and review before we’re able to turn our attention to the system that the authors designed. Each one of the words in the title is like a jigsaw puzzle piece. It won’t be until the end of the episode that the pieces start to come together, and the whole picture starts to make sense. So bare with me, it’s going to be a journey. Looking back up at the title now, I want to start in the middle. What’s an InSAR satellite? What do they do? Once we get a grasp of that, the problem-space will start to crystallize and we’ll have set the stage for the solutions to those problems...
Request this Episode
Release Date: February 7th 2025
The first thing you need to know is that there’s something called "Headway". Headway is the time it takes for a following-vehicle to reach the same point that a leading-vehicle just passed. It is typically measured in seconds and helps determine safe following distances in traffic. If you’ve ever heard someone say to "give x seconds to the car in front of you", that number of seconds is headway. Today’s research largely concerns highway driving, where headway is crucial. Too much headway, and the cars behind you will start passing you; leapfrogging to the space you should have been occupying. Too little headway and not only do you risk running into the back of the next car, but you’re much more likely to feel the need to hit the brakes hard when the next car’s brake lights come on. For your passenger, that results in jerkiness. Prior to this paper, a number of systems...
Request this Episode
Release Date: February 6th 2025
So what does it mean to analyze sounds? What does that process entail on a technical level? Well, sound first gets captured by a microphone as an analog waveform (a continuous variation of air pressure over time). This waveform is then converted into a digital format using an analog-to-digital converter (ADC), which samples the sound at a high frequency (e.g., 16 kHz or 44.1 kHz, etc). This process quantizes each sample into discrete numerical values. The resulting digital audio is typically stored in formats such as WAV, FLAC, or MP3, which preserve these numerical representations in various levels of compression and fidelity. A digital sound file is essentially just a sequence of numbers that encode the amplitude of the sound wave at each sampled point in time. For example, in a WAV file, which is an uncompressed...
Request this Episode
Release Date: February 5th 2025
The multidimensional knapsack problem is a combinatorial optimization problem that extends the classic knapsack problem by introducing multiple constraints. In its simplest form, the traditional knapsack problem involves selecting a subset of items, each with a given weight and value. You need to maximize the total value while ensuring that the weight does not exceed a fixed capacity. The multidimensional variant generalizes this by considering multiple resource constraints instead of a single weight limit. Each item consumes resources from multiple dimensions, and the total usage across all selected items must remain within the predefined capacity for each constraint. This problem arises in many real-world scenarios where limited resources must be allocated efficiently. In logistics, for example, a company may need to pack goods into containers, each subject to multiple restrictions such as...
Request this Episode
Release Date: February 4th 2025
As you would probably guess, they decided to use a convolutional neural network (CNN) as the classifier itself. This is a type of artificial neural network commonly optimized for analyzing visual data. Rather than train from scratch, they decided to use transfer learning, where a pre-trained CNN, specifically VGG16, was adapted to their new use-case. VGG16 is a well-established architecture originally trained on ImageNet. By leveraging the knowledge that’s already encoded in its convolutional layers, it can extract meaningful patterns from potato leaf images without requiring a vast dataset for training. All the researchers needed to do is replace the final classification layers of that model with custom layers suited for their specific classification task. To enhance interpretability, they incorporated Grad-CAM: Gradient-weighted Class Activation Mapping. This is a technique that...
Request this Episode
Release Date: February 3rd 2025
Where NewSQL diverges from NoSQL is in its treatment of consistency. NoSQL databases were developed to solve the scalability limitations of relational databases. Most of them accomplish this by relaxing strict consistency guarantees in favor of availability, partitioning, and performance. Rather than focusing on ACID, they often follow a model called BASE: Basically Available, Soft state, Eventually consistent. BASE allows nodes in a distributed system to return slightly outdated data in exchange for lower latency and higher throughput. While this works well for applications like social media feeds and real-time analytics, it presents serious challenges for use cases that require precise, immediate consistency. NewSQL avoids these pitfalls by maintaining ACID compliance, ensuring that distributed transactions are just as reliable as those in a traditional relational...
Watch the Trailer
Release Date: February 2nd 2025
They decided to use a combination of two different technologies: SARIMA and LSTM. They would build, train, and test both independently, and then consider if they should be ensembled together. Let's learn a little bit more about these two model types. SARIMA, or Seasonal AutoRegressive Integrated Moving Average, is an extension of the ARIMA model that has been modified to handle seasonal time series data. The AutoRegressive (AR) component models a time series as a linear combination of its past values, using lagged observations as predictors. The Integrated (I) component applies differencing to remove trends and make the series stationary. The Moving Average (MA) component captures dependencies on past forecast errors, adjusting predictions based on previously observed deviations. SARIMA extends this by adding seasonal autoregressive (SAR), seasonal moving average (SMA), and...
Watch the Trailer
Release Date: February 1st 2025
The process begins by inserting each tuple into the tree. The root node contains all data tuples, and from there, branching occurs based on the presence or absence of attributes. Intermediate nodes store weight-based decisions rather than attribute values, ensuring that tuples with missing values are assigned correctly. The tree grows downward, with leaf nodes representing fully classified groups of tuples that share identical missingness patterns. Once classification is complete, the tree’s indexing mechanism ensures that data can be efficiently retrieved and processed. Since each leaf node corresponds to a distinct missing data pattern, skyline queries can be executed within each bucket separately, minimizing unnecessary dominance comparisons across tuples with incompatible attribute sets. One of the other key innovations here is the introduction of Optimal Virtual Points...
Request this Episode
Release Date: January 31st 2025
It's called the Artificial General Intelligence Modularity Hypothesis. I’m just going to call it “Modularity” for short. The Modularity argument goes like this. Two premises and a conclusion: Premise 1: The human brain is not one big thing. It’s separated into regions. There’s the occipital lobe that processes vision, and the amygdala that regulates emotion. There’s the frontal lobe, the prefrontal cortex, the cerebellum, and many others. No single one of these regions is a brain unto itself, they each have a narrow focus. But collectively, they are more than the sum of their parts. Intelligence, as we know it, is an emergent property that mysteriously manifests itself only when these disparate components connect, interact, and communicate. Premise 2: Each AI sub-field (and each new type of model) is equivalent to a different region in the brain. While LLMs are highly sophisticated, they’re really just...
Watch the Trailer
Release Date: January 30th 2025
If you’re operating a website at scale, then it’s very unlikely that you’re running it on a single server. When you were setting up your DNS, you probably didn’t have your A-Records point right to the static IP of a server. Instead, you most likely have your A-Records set to an IP address that maps (either statically or ephemerally), to a load balancer. This load balancer (or a set of load balancers), sits in front of your cluster of servers. And when traffic comes in, the load balancer picks which server should actually receive the traffic and process the request. I make a request to your website, and it gets processed by Server-1. someone else makes a request and it gets processed by Server-2, etc. This allows your system to scale horizontally. Your cluster of servers is able to divide up the work that would have otherwise all fallen on one of them. These kinds of balancers are...
Watch the Trailer
Release Date: January 29th 2025
The Traveling Salesman Problem (TSP) is a classic optimization challenge. There’s a set of points (locations), and a “salesman” (or in this case, a delivery vehicle) who must visit each point once, and then return to the starting point. And he must do all this in the shortest total distance. On the surface, it seems fairly straightforward. You might assume that with enough computational power, you could just check all the options fairly quickly. But you can’t. The complexity of TSP grows exponentially as the number of points increases. The number of total possibilities is the factorial of one-less than the number of points. So with just a handful of destinations, it’s feasible to calculate every possible route and choose the shortest one. But, as the number of points grows, the solutions can’t be built incrementally, so there is a combinatorial explosion. This is what we call NP-Hard...
Watch the Trailer
Release Date: January 28th 2025
Today’s paper is about advancements in an algorithm called k-Nearest Neighbors. If you’ve listened to some of our Machine-Learning episodes, you’ve undoubtedly heard that term before. k-Nearest Neighbors (or kNN for short) is everywhere, and is used for everything. Why? Because it’s simple and it's effective. It’s not necessarily always the best tool for the job (and later we’re going to see how these authors are trying to modify it to be better at more things), but it is pretty good at a lot of use-cases. That’s why you’ll see kNNs used in ecology, in linguistics, in sports analytics, and even in archaeology. So before we get into this paper, let’s talk about kNN for a while. How does it work? What do “distances in feature space” even mean? And why is this algorithm so good at calculating them? Let’s get into it....
Watch the Trailer
Release Date: January 27th 2025
Our paper today looks at how one country, Bangladesh, is using data-science to avoid repeating the mistakes of the past. This research is part of a larger initiative they’re using to equip the agricultural sector with the statistical and computational tools they need to adapt their growing strategies to climate change. In this paper they explore the relationship of temperature to rice yields. The authors use statistical analysis methods to discover trends hidden in daily temperature data over time, and show how they impact crop growth. We’re going to walk through how exactly they’re doing that. We’ll look at descriptive statistics, the Man-Kendall Test, and Wavelet Coherence. We’ll show how these researchers use those tools to get a multifaceted understanding of the variables that are critical to drought intervention. When it established itself as an independent...
Watch the Trailer
Release Date: January 26th 2025
RETTS was developed in Sweden in the early 2000s, and is a structured and evidence-based triage model designed for high-resource healthcare settings. RETTS categorizes patients based on a combination of both vital signs and clinical symptoms. When they arrive at the ER, patients undergo an initial assessment of a few key physiological parameters (like heart rate, respiratory rate, blood pressure, oxygen saturation, and temperature). These values are then cross-referenced with a rubric called Emergency Symptoms and Signs (ESS). ESS provides guidelines for interpreting the clinical context of the patient's presentation. The system assigns patients into one of five urgency levels. Each urgency level is designated as a specific color, and is mapped to a recommended maximum waiting time. There is: Red (immediate care), Orange (20 minutes), Yellow...
Request this Episode
Release Date: January 25th 2025
Traditional 3D reconstruction techniques often rely on explicit geometric representations such as point clouds, meshes, or voxel grids. This new idea they presented: NeRF (Neural Radiance Fields) represents a scene implicitly using a neural network that encodes the complex interplay of light, materials, and spatial structure in a continuous function. This function maps a 3D spatial coordinate and a 2D viewing direction to corresponding color and density values, effectively capturing the scene’s volumetric properties in a highly compact form. By leveraging the neural network’s ability to learn complex patterns, NeRF can reconstruct intricate scene details and realistic lighting effects that traditional methods struggle to achieve. At the core of NeRF's rendering process is a technique called volume rendering, which allows for the synthesis of...
Request this Episode
Release Date: January 24th 2025
While TCP is a connection-oriented protocol, QUIC is a UDP-based protocol that combines transport and application-layer functions for faster performance. It works by multiplexing multiple streams over a single connection and reducing handshake overhead. It supports connection migration, enables 0-RTT handshakes, provides built-in congestion control, multiplexes streams without head-of-line blocking, and minimizes latency. In other words: QUIC establishes connections quickly, handles multiple streams efficiently, and adapts to changing network conditions without requiring multiple round trips for handshakes. So yeah, it’s really fast. But, it also takes a novel approach to security. While QUIC is a transport-layer protocol, it includes features that are traditionally found in both transport and security layers. So essentially it’s doing the work of something like TCP plus TLS, except that...
Watch the Trailer
Release Date: January 23rd 2025
There are two broad categories of CRDTs: state-based and operation-based. In state-based CRDTs, each replica periodically shares its entire state with others, and a merge function is used to combine different states into a unified, conflict-free version. The merge function is designed to be idempotent (meaning that applying it multiple times has the same effect as applying it once), associative (meaning that the order of merging does not matter), and commutative (merging in different sequences leads to the same result). Operation-based CRDTs, in contrast, rely on propagating individual operations rather than entire states. Each operation is designed to be deterministic and to produce the same effect across all replicas, provided that it is delivered in causal order. This reduces the amount of data that needs to be transmitted. So in practice, operation-based CRDTs are...
Watch the Trailer
Release Date: January 22nd 2025
DNA is unbelievably compact and efficient at storing and encoding information. As a medium, it can be stored without electricity (in dry pellet form or dissolved in a solution), for thousands and thousands of years. We’ve understood how DNA stores information since Watson & Crick discovered the double-helix in the ‘50s. But until recently we’ve only been able to marvel at that storage mechanism from afar. We haven’t been able to use DNA to store information ourselves. The storage mechanisms we’ve come-up with instead are, well…way worse. Hard drives, flash drives, magnetic tape, CDs, floppy disks, etc, don’t come anywhere close to DNA’s efficiency or durability. In order to use DNA as a storage medium, we’d need to be able to read and write DNA. That is “sequence” and "synthesize" it. Both of these processes have historically been cost prohibitive, but that’s changing...
Watch the Trailer
Release Date: January 21st 2025
The ViTs process images as a sequence of non-overlapping patches rather than pixel arrays. Unlike CNNs, which rely on local receptive fields and weight-sharing, ViTs divide an image up, and then convert the pieces into flattened feature vectors. These vectors are then passed through an embedding layer that encodes positional information to maintain spatial context. The core of the architecture is its multi-head self-attention mechanism, which allows the model to attend to different regions of the image simultaneously. Since it’s a transformer, it’s able to capture complex relationships between distant regions of the tissue sample that may not be immediately adjacent but are diagnostically significant. Each transformer layer applies self-attention operations followed by normalization and feed-forward layers to refine the feature representation....
Watch the Trailer
Release Date: January 20th 2025
On March 23rd, 2016, Microsoft released a product that...well...they’re probably hoping we’ve all forgotten about by now. It was called “Tay”, and it was a Twitter-bot using the handle @TayAndYou. It was designed to interact with users and learn from conversations, in real-time. People would tweet at Tay, Tay would tweet back, and, the hope was, that it would become a full-fledged thoughtful netizen, carrying on interesting conversations and showcasing the advancements that Microsoft had been making in AI. It’s difficult to describe just how quickly this project went off the rails. Within hours, Tay began to mirror and amplify the language patterns it encountered, generating increasingly offensive statements. As users trolled the bot, deliberately feeding it inflammatory and discriminatory text, Tay’s learning mechanisms, which lacked strong ethical or content moderation filters, quickly adapted to replicate those inputs...
Watch the Trailer
Release Date: January 19th 2025
What is Wasm today? It’s changed a bit from its roots, but it’s kept the core idea unchanged: developers can write code in languages like C, C++, Rust, (or any one of many other languages that now have a Wasm compilation target), and can compile their code and ship it to anywhere that Wasm is supported. Despite the name, Web Assembly actually escaped the confines of the browser years ago. It is now just a first-class runtime for a number of different environments. Wasm runs in the cloud, it runs on the edge, it runs on iOT devices, it runs on embedded systems and gaming engines. It’s everywhere. And as Wasm has expanded the scope of its deployments, it’s also begun to run into a new set of problems. The kind of problems that weren’t necessarily issues back when it was confined to the browser. Today’s paper is focused on just one of those issues. Just one of the growing-pains that...
Watch the Trailer
Release Date: January 18th 2025
What do bees, ants, moths, bats, fish, birds and whales have in common? Give up? They all have optimization algorithms named in their honor. And today’s article is focusing on a new variant of that last one: whale optimization. In this paper it’s being used to solve a mapping problem for something called a Digital Twin Network (DTN). But the authors didn’t just solve that problem the one way, no. They actually solved it seven different ways, with seven algorithms, and then benchmarked them against each other. That gives us the opportunity to spend today’s episode taking a stroll through algorithm land. We’re going to ground ourselves in the problem-space very briefly, then spend the rest of the time learning about seven different algorithms. How they work, their strengths and weaknesses, and which one ended up being the best fit for the DTN problem. Let’s jump into it...
Watch the Trailer
Release Date: January 17th 2025
Have you ever wondered where weather forecasts come from, or how meteorologists generate them? Well, it’s a multi-step process, and there are a lot of moving pieces at play. But one of the most important parts is something called numerical weather prediction (NWP). NWP is the process of taking what’s called an “initial condition” and applying mathematical and physical equations to it, to turn it into a forecast. If the initial condition is correct, then the forecast has a good chance of accurately predicting future weather conditions. If the initial condition is wrong, then the forecast will likely diverge from reality. So, what is that initial condition, and where does it come from? The initial condition is the best estimate of the current state of the atmosphere...
Watch the Trailer
Release Date: January 16th 2025
Imagine that you went out and bought a big jigsaw puzzle. You take it home, open the box, flip it over and dump all the thousands of little pieces on the table. What’s the next thing you do? Well, I’m no puzzle-master, but the next thing I’d do is categorize all the pieces. This is basically a “classification” task. I have to look at each one, and say “that one is sky, that’s grass, that’s a piece of a building, that’s a part of a car, that’s a part of an animal, that’s a texture, that’s a pattern” etc. If I was a computer-vision model, and I was looking at pixels instead of puzzle-pieces, then this task would be called “semantic segmentation”. In semantic segmentation, a model scans through an image and classifies each pixel into a predefined category. As it reaches each pixel, it analyzes its features, and determines its class based on learned patterns...
Watch the Trailer
Release Date: January 15th 2025
In order to understand today’s paper, we need to start by wrapping our heads around the problem-space it’s working in. AMR: Automatic Modulation Recognition. Imagine there's a war-zone. And in that war-zone, two armies are duking it out: the red army and the green army. The red army has soldiers scattered across the land, and these soldiers carry devices that let them communicate with each other. (Radios, satellite phones, infrared beacons, field tablets etc). These devices all communicate somewhere on the spectrum, and there’s a finite amount of spectrum available to use because there are only so many bands of electromagnetic radiation. So, the issue is, the green army has access to all the same spectral bands as the red army. So if the red army soldiers try to send messages to each other across these frequencies...
Watch the Trailer
Release Date: January 14th 2025
The planning and execution module decomposes tasks into executable sub-tasks, again using LLM-based reasoning. Each task is processed into a sequence of commands. These commands are specific to whatever kind of robot the system is running onboard. And they could either be straight from the SDK or from a source we haven’t mentioned yet. The system has an onboard list of previously acquired skills. They call it the skills library. We’ll come back to how skills get into the library in a moment. For now, just know that as the robot is exploring its environment, the actions it previously tried that were successful in other contexts are available to it in the skills library. Once a series of actions is chosen, the execution module generates code to control the robot. During execution, the module ensures task...
Watch the Trailer
Release Date: January 13th 2025
When the algorithm gets called you need to pass it an incidence matrix, which is just the mathematical version of the diagram I described earlier. It starts by just initializing an empty list. This is where it’s going to store any uncovered transitions that it finds. Next, it flips (transposes) the rows and columns of the incidence matrix. This is just to prepare it for the next steps, which involve analyzing the transitions mathematically. Now, the algorithm reworks the transposed matrix into something called Reduced Row Echelon Form (RREF). This is a cleaned-up version where patterns in the data become much clearer. It involves reorganizing and simplifying the rows to highlight key features. With the RREF in hand, it then looks at each row (representing places in the system) to find transitions that...
Watch the Trailer
Release Date: January 12th 2025
There are two main approaches for this: one-step classification, and two-step classification. One-Step Classification: A whole image classifier (WIC) processes the entire image (or frame), including the entire background, and makes predictions based on all visible elements in the photo, without isolating the specific object of interest. Two-Step Classification: In this approach an object detector attempts to identify any prominent object in the frame, and it returns the coordinates of the bounding-box that encompasses that object. On the second step, an image classifier looks at the image, but only looks within the bounding box (the region of interest) determined in the first step. Both of these approaches work, and both are quite popular. But the ML community has been split over which option is better. Today’s paper is from a research team at Google, and they’re trying to put an end to that debate...
Watch the Trailer
Release Date: January 11th 2025
Bucket 1 is the Bloaters. These arise when something in your codebase has grown too large to handle gracefully. These smells don’t immediately break your system but instead slowly erode its maintainability. Examples include the “Long Method”, where a single function becomes a sprawling labyrinth of logic. And the “Large Class”, which is stuffed with so many responsibilities it’s hard to tell what its primary purpose is. Another classic bloater is the “Long Parameter List”, which is a sign of poor abstraction, and requires the caller to juggle too many pieces of data at once. Left unchecked, bloaters bog down your code, making it harder to navigate, extend, or debug. Bucket 2, the Object-Orientation Abusers, is all about misusing or misunderstanding fundamental principles of OOP. This bucket includes...
Watch the Trailer
Release Date: January 10th 2025
The year was 1985, and Ross Freeman had a problem. He was trying to design hardware solutions (computer chips) that could be quickly adapted to evolving requirements. But the existing options of the day didn’t fit. General-purpose microprocessors, ASICs, and hardwired logic boards weren’t flexible enough. Programmable logic boards and devices weren’t powerful enough. He needed a platform that was as capable as an ASIC, but could be reprogrammed after manufacturing, without the cost and time of creating a new chip for every use case. He envisioned a new kind of device that combined the efficiency and performance of hardware, with the adaptability of software. A computer chip that could be iterated upon; modified and customized over and over again. The result was the world’s first FPGA: the Field-Programmable Gate Array...
Watch the Trailer
Release Date: January 9th 2025
At a high level, you can think of their system as a pipeline. It leverages both multi-temporal data and advanced attention mechanisms. At its core, the model integrates a combination of spatial, spectral, and contextual information. That is all just a fancy way of saying two things: It looks at different photos from the same area over time. Clouds move, so if you pay attention to one big area, and take a series of photos over time, the areas that are cloud-free will change over time. And if you collect enough photos eventually you’ll have at least one cloud-free view of every region in that area. Then you’ll be able to collage together a single photo that is wholly cloud-free. Snow-cover and cloud-cover may look the same in photos, but that’s only representing the visible light spectrum...
Watch the Trailer
Release Date: January 8th 2025
There are two distinct machine learning tasks in play here: embryo-ranking, and implantation-prediction. They are similar-enough problems that clinicians use the same machine learning model for both tasks. The authors are saying that these two activities actually have slightly conflicting objectives, and therefore using the same model for both is wrong. The status-quo is fine for implantation-prediction, but when you use that model for embryo-ranking, the rankings are incorrect. This is because of a very subtle and difficult-to-understand phenomenon that we have never talked about on Journal Club before. It’s called the “shortcut learning” bias. And in the case of these IVF procedures, shortcut-learning is sabotaging the patient’s chance at a successful pregnancy....
Watch the Trailer
Release Date: January 7th 2025
Imagine you’re writing a function called “redact_names”. It takes one parameter, a string. Your function needs to accept the string, replace any names with a series of asterisks, and return the string. It needs to handle all the corner-cases I mentioned above (full names, partials names, and initials). But remember, people’s names may have any kind of capitalization or no capitalization at all. They may be one word, or one letter, or a series of words and letters. A name could be two words, or three, or four or more. There may even be standalone sentences that contain nothing but a name. And in other cases there may not be any names in the passage at all. Think that’s not complicated enough? It gets worse. Medical writing is often chock-full of what are called “medical eponyms”...
Request this Episode
Release Date: January 6th 2025
There’s been a lot of talk about GPUs lately, and almost all of that talk has either been about graphics processing, crypto mining, or machine learning. But, there is a small community of developers who are thinking beyond those buckets. Developers who are wondering out loud: "What if we wrote lots of programs to run directly on GPUs? What if we used their parallel structure to handle the concurrency problems inherent in all kinds of applications?" This idea is broadly referred to as "GPGPU". It stands for General Purpose computing on GPUs. Today we’re going to look at a paper that embraces the GPGPU paradigm. The authors are building a SLAM system (a Simultaneous Localization and Mapping tool), that allows a robot to map their environment and understand their position at the same time...
Watch the Trailer
Release Date: January 5th 2025
User-space ransomware works within the permissions granted to standard user accounts. This means it cannot modify system-level configurations or access kernel data. This also means, critically, that if you have backups or restore-points in protected areas that require administrator privileges, those are also safe. What the ransomware can do is encrypt most user files, as these are typically within the permissions of a standard account. Root-privileged ransomware, on the other hand, is significantly harder to detect and mitigate. Root-level ransomware can disable security tools, encrypt backups, and render entire systems inoperable. Today we're looking at a paper that presents a new mitigation option for user-space ransomware only. It won't really apply to anything operating with root privileges....
Watch the Trailer
Release Date: January 4th 2025
This story takes place at the Culham Centre for Fusion Energy (CCFE), located on the outskirts of Oxford in the UK. There are a number of different types of nuclear fusion reactors in the world, and the specific kind they’re working with at CCFE is called a tokamak. It’s originally a Russian design, and the name comes from an acronym that only makes sense in Russian, but roughly translates to "toroidal chamber with magnetic coils". If you’ve ever seen a photo of a fusion reactor that looked like a giant metal donut, that was probably a tokamak. The CCFE, like virtually all other nuclear-fusion sites, isn't trying to generate power for commercial purposes. Not yet, anyway. It’s a research facility. Before a fusion power plant can ever...
Watch the Trailer
Release Date: January 3rd 2025
At first glance, this paper appears to be about magnesium alloys. But if you go just beneath the surface, it’s really about LLMs, and the techniques you can use to squeeze more functionality out of them. In this paper the authors are trying to build an LLM that can function as an expert system for magnesium alloy phase-diagrams. That topic is extremely narrow, but incredibly deep. And in order to function as an expert, their model must incorporate a ton of highly specific information about that field, including thermodynamic principles, multi-component alloy systems, phase transition behaviors, and huge industry-specific datasets. They decided to try out three different methods to accomplish this: 1) Base (unmodified) LLMs with prompt-engineering. 2) Base LLMs with RAG. 3) New models created with SFT. SFT stands for Supervised Fine Tuning. In the last-year or so, OpenAI made an endpoint available where you can...
Watch the Trailer
Release Date: January 2nd 2025
If you’ve gotten used to creating VPCs on AWS or GCP, (or VNets on Azure), then you’ve probably become pretty spoiled. All those configuration options, and knobs and dials. The CIDR blocks and subnets and masks. All that comes out of the box when you’re working in the cloud. But if you’re administering a traditional corporate network for an enterprise, the picture is probably very different. These kinds of legacy systems don’t provide nearly the functionality that you might want and need to support the company’s growing demands. And that’s where this paper comes in. The authors are proposing what they’re calling an ENT: An Enhanced Network Topology that makes a number of changes all-at-once to quickly modernize older networks....
Watch the Trailer
Release Date: January 1st 2025
When you’re playing a slot-machine, you’re making a series of choices based on the information and feedback you’re receiving from the interface. Each pull of the lever represents a decision to act, based on the outcome of previous plays. Over time, you start to gather information about potential rewards, and you must decide whether to continue playing, stop, or change strategies. You are effectively balancing the uncertainty of future rewards against the results you’ve been observing. This is called a sequential decision-making problem, and it is very similar to many other decisions that you, (or another person, or a business, or an autonomous-car) need to make all the time. In many aspects of life, you're taking actions, observing outcomes, and making choices that (ideally) optimize your results over time....
Watch the Trailer
Release Date: December 31st, 2024
Wavelets are mathematical functions that are designed to analyze data at different scales or resolutions. Think of a camera with different zoom-levels or lenses. When you zoom in you see more detail but less of the whole context, and when you zoom out you see the whole picture, but fewer details. That's like a wavelet adjusting its focus to capture both the fine details or the broader trends of a signal. A wavelet transform uses these wavelets to break a signal into its components, separating coarse, low-frequency information (the "big picture") from finer, high-frequency details (the "small nuances"). This decomposition allows for efficient analysis while still preserving the essential characteristics of the signal. These authors use two types of transforms...
Watch the Trailer
Release Date: December 30th, 2024
At Journal Club, we spend a lot of time talking about technologies that are, from an environmental perspective, inherently unsustainable. At least in their current forms. It’s worth taking a moment to remind ourselves that a significant number of the GPUs that were mining crypto a few years ago, are still running at full throttle. It’s just that now they’ve been repurposed to power ML instead. So, all the same environmental critiques that were made of crypto back then could be made of AI today. One would hope that, today, all that energy is being put to a more noble use, but still...carbon footprint is carbon footprint. And I’m sure there's room for all of us to do better. Since we’re about to put a bow on 2024, I figured that now would be a good time to take a look at a few technologies that are on the other side of the fence. The green side...
Watch the Trailer
Release Date: December 29th, 2024
With all the fanfare around tools like ChatGPT, Claude and Gemini, it can be easy to think that transformer architectures are the be-all end-all of AI. That they represent the latest and greatest technology for processing text, media, and documents of all kinds. In reality, they don’t. There was a single point in time when transformer architecture was the bleeding-edge for the use cases it was designed to fill, but that was a few years ago. In the time since then, the research hasn’t stopped. And in fact, a whole new generation of models and architectures have been developed in recent years with the deliberate aim of overcoming some of the specific limitations of transformers. One of those new architectures is called Mamba...
Watch the Trailer
Release Date: December 28th, 2024
Some diseases are pretty straightforward to diagnose. If you're unfortunate enough to have something like hep-b or mononucleosis, your doctor should be able to order a test that will give you a definitive result fairly quickly. But that's not how all (or even most) diseases work. For a great number of illnesses: like chronic kidney disease, autoimmune disorders, or Wilson's Disease, there may not be an antibody test, or reactive test, or any other kind of rapid diagnostic kit. There's no single point-in-time way to tell if you have the condition or not. The only way your doctors can diagnose those kinds of illnesses...
Watch the Trailer
Release Date: December 27th, 2024
Let’s start at the beginning. Up until about two decades ago, you didn’t download software. You bought it at a store, in a box. It was a physical product. For that to happen, the software had to be put on some kind of medium, boxed, wrapped, put in a bigger box with others smaller boxes, then put on a pallet with other bigger boxes, wrapped in shrinkwrap, stored in a warehouse, sent to a distributor, trucked to the store, unpacked, and put on a shelf. There was a physical supply chain, and getting your code from your workstation to the customer’s machine took a very long time, and was very expensive. For this reason, software updates were...
Watch the Trailer
Release Date: December 26th, 2024
As Engineers we rarely think about infrastructure costs. But once an application reaches a certain scale (or when a startup runs out of free credits) we no longer have a choice. We have to think about it. We need to be able to predict financial costs just as accurately as we can predict uptime and load speeds. Off the shelf calculators (like the kind that the cloud-host provides) are fine for small things, but if you need to predict the costs of dozens or hundreds of interconnected pieces of infrastructure, you need something more robust. You need a financial model. The question is, how do you do that? How do you build one of those? That’s what today’s paper is all about...
Watch the Trailer
Release Date: December 25th, 2024
Today is a special episode of Journal Club. This is the first installment of a series I’m calling "Post Quantum Readiness". I’m going to sprinkle these episodes in every once in a while over the next year, or two, or however long they’re needed. While the goal of every Journal Club episode is to get you up to speed on some new research (and this episode is no exception in that regard), this "Post Quantum Readiness" series is going to be much more focused on practical advice. We’re going to be looking at research papers of course, but for this series we’re specifically trying to extract three things from them: 1. What you need to know now. 2. What you’ll need to know tomorrow. 3. Calls to action. Steps you can take immediately to start to get ahead of this....
Watch the Trailer
Release Date: December 24th, 2024
Did you ever watch that show CSI: Crime Scene Investigation? In many episodes, there will be one character sitting at a computer pulling up video surveillance of the crime scene or somewhere nearby. And another character will see some blurry suspect or object in the video and say something like “Wait, pause that. Now zoom in and enhance!” The person at the computer will type a few keystrokes, and just like magic, the grainy frame of the perpetrator will turn into a high resolution photo. This has been a Hollywood trope for decades. In reality, we are just now (in the last few years) getting to the point that we have the ability to do anything close to that with normal hardware. And that genre of technology is called SISR: Single Image Super Resolution...
Watch the Trailer
Release Date: December 23rd, 2024
An FNN is built using fractal blocks, which are modular units that repeat. Each block contains standard components like convolutional layers for extracting features from input data, and pooling layers for reducing dimensionality and retaining important information. What makes FNNs unique is how these blocks are organized: instead of stacking layers linearly as in traditional neural networks, the architecture branches recursively. At each recursive step, a fractal block is expanded into multiple sub-blocks, which are connected in parallel paths. But unlike in a true fractal, this only repeats for a range, not infinitely. Let’s walk through the points at which an FNN diverges from a CNN, and what it does differently. In a traditional CNN, layers are stacked, (as I mentioned), sequentially. With each layer processing...
Watch the Trailer
Release Date: December 22nd, 2024
Optical computers are fundamentally different from traditional electronic computers. In essence, they use photons (light) rather than electrons to process and transmit information. This shift in medium leads to several critical distinctions in how components operate and interact. Recall that at the heart of traditional computers is the transistor, a semiconductor device that controls the flow of electrons to perform logical operations and amplify signals. In contrast, optical computers leverage optical transistors (or similar photonic components) that control and manipulate light. Unlike electronic transistors, which rely on voltage or current, optical transistors operate through phenomena such as nonlinear optical effects, where one light beam can modulate another. This enables optical components to switch signals at speeds closer to the speed of light, significantly faster than the electron mobility in silicon...
Watch the Trailer
Release Date: December 21st, 2024
A GAN is a type of machine learning model that is able to generate (synthesize) new data that resembles its training data. It doesn’t just take one piece of its training data and give it to you, it creates new data that was not actually part of its training, but that shares a number of similar characteristics with the data it saw during training. It does this by having two different core components: a generator, and a discriminator. Think of the generator as the artist, and the discriminator as the critic. The generator creates new things and tries to pass them off as being part of the original training set. The critic evaluates them, either saying "no, I can tell this wouldn't have been in the training set", or "yes, this fooled me. It looks like the training data". The process of synthesizing data in a GAN can be thought of as a back-and-forth conversation between these two “adversaries”. Thus the name: Adversarial Network....
Watch the Trailer
Release Date: December 20th, 2024
Your mission is to travel to a new site and install sensors, monitoring, and edge nodes as usual. But there’s a catch. This cultural heritage site… is underwater. It’s not a dig-site or a cave, it’s an ancient shipwreck. So now you’ve got some thinking to do. How does the heritage-site being underwater change what you need to build? What’s the same, and what’s different? Let’s walk through it. There are five key differences: 1. There’s no power in the middle of the ocean. 2. Radio signals attenuate in water. 3. GPS is ineffective under water. 4. Not only is everything going to get wet, but pressure is high, temps are cold and saltwater corrodes everything. 5. Connectivity to the cloud will be possible, but limited. You’re no longer working in vanilla IoT. This is IoUT, the Internet of Underwater Things....
Watch the Trailer
Release Date: December 19th, 2024
The authors argue that traditionally, you would solve this problem by gathering all your data then solving it as a fuzzy linear programming (FLP) problem. But, in this case, they say that’s not good enough. So this paper is introducing a new strategy that gets better results. It’s a five-phase combinatorial method that integrates five separate ideas: Triangular Intuitionistic Fuzzy Numbers (TIFN), Intuitionistic Fuzzy Linear Programming (IFLP), Realistic Robust Programming (RRP), Chance-Constrained Programming (CCP) and Augmented Epsilon Constraint (AUGMECON). Their argument is that this new system improves on FLP by allowing hesitation, ensuring robustness, considering both satisfaction and non-satisfaction levels, and generating Pareto-optimal solutions....
Watch the Trailer
Release Date: December 18th, 2024
If I say peanut butter and _______, you’ll probably be able to guess “jelly”. That's just something your brain can do. In computer science we’d call this predictive modeling. That’s been the foundation for a lot of the things we use everyday, from the autocomplete in the search bar, to ChatGPT. In today’s paper we’re going to see predictive modeling applied to data compression. The authors have developed a system that compresses files to a smaller size (than normal compression-algorithms) by just leaving out chunks of the data. The idea being: they’ll be able to guess what was removed when they go to decompress it. While this approach isn’t fundamentally different from how modern compression works (other algorithms do use predictive modeling), these authors introduce new mechanisms, and error-correction systems that are genuinely novel. They call their new system: COMPROMISE....
Watch the Trailer
Release Date: December 17th, 2024
Today we’re going to be looking at a system-design paper that is not actually for a standalone application at all. The system the authors are building is actually just a very complex plugin (or extension) for CRMs: Customer Relationship Management platforms. Specifically, this one works with Salesforce, and the "Apex" developer platform. That being said, the Salesforce-specific code is really just a tiny part of it. The vast majority of their work is around training an ML model that can review an incoming lead, and attempt to “qualify” that lead programmatically. It then assigns that lead a score, which gets pulled into the CRM...
Request this Episode
Release Date: December 16th, 2024
If you open up the PDF for today’s article, you’ll see that the meat of the paper is all formulas. It’s just equation after equation. I’m not going to read them to you, that would be an awful episode. Instead, I’m going to do my best to break down the mathematical foundation on which this paper is based. Then I’ll explain how they’re using those formulas, and what they’re trying to accomplish. The math in this paper is all based on Set Theory, so we’re going to start there. Technically, it’s all based on single-valued neutrosophic soft sets…but we’ll need to build-up to that. Set theory is a foundational area of mathematics that deals with the study of collections of objects, known as sets. It provides a framework for defining and manipulating collections of elements. In set theory, operations like union, intersection, and complement define how sets relate to one another, while concepts like...
Request this Episode
Release Date: December 15th, 2024
If Section 230 gets nerfed by the incoming administration, social media sites will need to get a handle on hate-speech and extremism very quickly. Every modern social media platform has a hate-speech problem, and this is nothing new. In fact, the authors of today’s paper were able to find 110 different studies in which the investigators attempted to build a system for identifying and classifying that kind of language and behavior. Then, they ran a "systematic-review" (which is like a fancy meta-analysis), on all the papers to determine the best way that we can use technology to identify ideological extremism. That’s what this paper is about. It’s a head to head shootout between six different approaches, to figure out which is most effective. Let’s review each type of approach, then look at how these authors conducted their comparisons, and finally take a look at their conclusions....
Request this Episode
Release Date: December 14th, 2024
In the 1970’s and 80’s, a strange fish was regularly sighted off the coast of Miami, bobbing up and down in the water. It was big, brown, and oddly...rectangular. The locals had a name for it: the “square grouper”. The square grouper was, of course, not actually a fish. It was a compressed bale of illegal drugs that traffickers had dropped out of a plane into the water so that a smaller vessel (often a speedboat) could scoop it up. In modern times, this is known as the “drop-off” method, and in today’s paper the authors attempt to train a model that can detect when a boat is engaging in this kind of illegal behavior. We’re going to walk through what the modern-day version of a drop-off looks like, how other researchers have tried to detect this behavior before, and exactly what these authors are doing differently....
Request this Episode
Release Date: December 13th, 2024
The complexity came mostly in step 2: Getting low-power edge nodes to run lightweight models. Deploying models on edge devices like NVIDIA Jetson Nano, which have limited GPU cores and memory, is challenging. Modern deep learning and CV architectures often involve millions or even billions of parameters, and they need to perform 3D convolutional operations to extract spatial and temporal features from video frames. This is necessary for accurate action recognition, but it requires substantial memory and processing capabilities. To make this possible in a resource-constrained environment, the authors applied Knowledge Distillation, a model compression technique that creates smaller, more efficient models while preserving much of the predictive accuracy of larger, more complex models. In this process...
Request this Episode
Release Date: December 12th, 2024
The routing algorithm starts by decoding the route polyline string (received from Google Maps), into a series of waypoints. Each waypoint is a latitude-longitude pair. The waypoints are then filtered to reduce unnecessary computational overhead. A filtering service selects only waypoints that are at least 500 meters apart, minimizing the total number of points while maintaining path accuracy. The immediate-waypoint service then takes in that information, and identifies the next two waypoints along the path from the current position of the emergency vehicle. These waypoints guide the real-time navigation process and serve as reference points for subsequent route recalculations. The real-time location of the vehicle is continuously tracked, and the route is continuously recalculated in real-time....
Watch the Trailer
Release Date: December 11th, 2024
Back when I was managing Engineering teams, code coverage was everything. I spent an inordinate amount of time pushing our Engineers to raise the coverage-level as high as we could raise it. More unit tests, more component tests, more end-to-end and integration tests, more smoke tests, more everything. For me, the coverage-reporter was the strongest indicator of our code quality and the stability of our system. Statements, branches, functions, lines: those breakouts were my scorecard, and my teams probably spent more time improving those numbers than anything else. Back then, if you would have given me the option of reducing the number of tests on our codebase, I never would have taken it. In fact, I probably would have laughed in your face. But...The Times They Are A-Changin...
Watch the Trailer
Release Date: December 10th, 2024
Cuckoo Search Optimization (CSO) is a metaheuristic algorithm inspired by the unique reproductive behavior of certain cuckoo bird species. These birds lay their eggs in the nests of other bird species, relying on the host birds to incubate and raise their offspring. If the host bird detects an alien egg, it may either abandon the nest or destroy the egg. This ensures that only the most deceptive and well-hidden eggs survive. This biological process serves as a metaphor in CSO, where "eggs" represent potential solutions to an optimization problem, and "nests" are different candidate solutions in the search space. The algorithm's ability to balance exploration and exploitation of the search space...
Watch the Trailer
Release Date: December 9th, 2024
Temporal Graph Networks (TGNs) extend GNNs by incorporating time-sensitive interactions into the graph structure. Unlike static graphs, which capture a frozen view of relationships, temporal graphs model how relationships evolve over time. TGNs process sequences of time-stamped events, making them well-suited for dynamic environments like online payment systems where transactions occur continuously. Each time a relevant event happens—such as a credit card being linked to a device or a payment being made—TGNs update their understanding of the network state. TGNs rely on event-based temporal graphs (ETGs) rather than snapshot-based graphs. Snapshot-based graphs divide...
Watch the Trailer
Release Date: December 8th, 2024
In ML, matrices are vital because they naturally encode data, models, and computations in a compact and efficient manner. You might have previously heard that ML runs on vectors, and that’s largely true, but matrices are just as important and often work in conjunction with vectors. Vectors represent individual data points or model parameters, while matrices enable large-scale data processing by grouping vectors together. So it’s really both of those concepts together that form the numerical underpinning of machine learning. Matrices support a wide range of tasks that ML pipelines rely on. Notably: data representation, model parameterization, and numerical computation....
Watch the Trailer
Release Date: December 7th, 2024
Why embark on any of this at all? What problem are they actually solving? What unmet need is there that they’re trying to fill? The first thing to know is that we're not talking about microservices running on big servers or the public cloud (AWS, GCP, etc.). We're talking about microservices running in highly resource-constrained environments, specifically IoT devices. This could be a smart thermostat, a wearable health tracker, a factory floor sensor, or a wildlife monitoring system. Either way, don't think of the playing field as a big hulky processors with a ton of horsepower and memory. It's the opposite of that....
Watch the Trailer
Release Date: December 6th, 2024
Every stack is a LIFO queue (last-in-first-out), so there’s only one solution, right? Wrong. It’s more complicated than that. Here’s why. And this is what makes it NP-Hard: Each container that is in the way of the container you want to reach is called a “blocking” container. Every time you move a blocking container, you have to decide where to put it, and that decision has consequences for future retrievals. Not only does each stack have height limitations, but the placement of that container on another stack can either make future retrievals easier or harder depending on the sequence of container requests that come later. If you move a container to a stack where it blocks another container that will be needed soon, you’ve effectively created an additional retrieval problem for yourself....
Watch the Trailer
Release Date: December 5th, 2024
If you just sent those 10 pieces as is, any eavesdropper could easily reconstruct the message. So you don’t do that. This is where that agreed-upon key comes in. You perform a mathematical operation called modular exponentiation on each of the 10 parts, using the key. Modular exponentiation involves raising a number to a certain power (the exponent) and then finding the remainder when divided by another number (the modulus). This operation is commonly used in cryptography because it's easy to perform in one direction but very difficult to reverse without knowing the exponent, which, in this case, is the key....
Watch the Trailer
Release Date: December 4th, 2024
The authors of this paper were looking for ways to improve the bootup (or “formation”) speed of a certain type of network between IoT devices. In a power-loss scenario, like a blackout these devices would need to boot back up and because of the complexity of the network they were establishing, that would take quite a long time. There’s just a lot of information and configuration that needs to be sent back and forth between different nodes, and this takes awhile. All these authors did, and all this paper is about is caching a little bit of that data. That’s the whole idea. Each node in the system needs...
Watch the Trailer
Release Date: December 3rd, 2024
Imagine it’s the early 2000s, and you’re working as a Software Engineer at one of the budding social networks of the time. It could be a Facebook, a MySpace, a Friendster, a Twitter, etc. What makes your work special is that you’re part of what’s increasingly being known as “Web 2.0.” Gone are the days of static websites and individual user experiences. This is the dawn of the dynamic web and shared experiences. Social experiences. Your company is not just storing data about users and their individual flows through the application, but also how each user is interacting with each other user. You have new concepts that describe this. For the first time, you have things like “following” and “follower.” You have “friends,” you have “groups” and social circles. You have relationships....
Watch the Trailer
Release Date: December 2nd, 2024
If you live in a region with an autonomous taxi pilot program, your local evening news is probably having a field-day with all the ridiculous (and often hilarious) mistakes that these cars are making. Some of these mistakes are issues with sensors, some are issues with training, and some are issues with law enforcement: like the viral video of the Cruise car in SF a couple years ago that got pulled over by the police and for some reason drove away during the traffic stop. But a lot of the issues lately have been stemming not from the robotaxis interacting with the environment, but actually from...
Watch the Trailer
Release Date: December 1st, 2024
The question these researchers are attempting to answer is can a Machine Learning model be trained to be able to look at the 3D point-cloud dataset and detect an aneurysm the way a medical professional could if they were looking at the 3D representation themselves. Based on my reading of the article I don’t think it’s their intention that such a model would replace what a Doctor would be doing, just putting an extra set of electronic eyes on the 3D-data to minimize the chance that a bulging area on a vessel would ever be missed. Luckily there is quite a lot of precedent when it comes to object-detection within a 3D space, so this research is really more of a shootout than an experiment....
Watch the Trailer
Release Date: November 30th, 2024
This paper is about a great piece of software with a terrible name. It’s called GloWord_biGRU. At a high-level this research is all about sentiment classification. It’s about building a tool that can classify user-comments, posts, and other freeform text as either positive or negative. Imagine that you work at a place like Youtube. You have lots of videos, people can thumb-up or thumb-down them to indicate if they like it, but they can also leave a comment. One popular video could have hundreds of thousands of comments, far too many for a team-member to review. But you do want to review those comments somehow. You want to feed the overall sentiment of each comment back into the system as if it was a thumbs-up or thumbs-down. So, you have a bunch of text, and you need to convert each blob into a boolean. True if it’s a positive comment, false if it’s negative. And of course you need to do this at scale, as quickly and cost-efficiently as possible. What do you do?...
Watch the Trailer
Release Date: November 29th, 2024
We all know that key-value stores are fast. They're fast at reading, they're fast at writing. The question is, could they be faster? Is it theoretically possible that there is still, to this day, a significant gap between their potential maximum, and their realized performance? Today's authors say yes, there is. In this paper they started with a host machine, and they analyzed the performance of the host's underlying storage hardware (NVMe SSD), to determine (at the lowest level) how fast data could actually be fetched-from and written-to disk. Then they setup a key-value store (RocksDB) on that host, then benchmarked reads and writes of that DB against the host's raw performance. Now what you'd expect to see is a little performance hit, explainable by the presence of the database application itself. It needs to run business logic, and that logic takes a little time. Understandable. But what the authors actually found was a far more significant...
Watch the Trailer
Release Date: November 28th, 2024
In this paper, the authors develop a detection and analysis framework, specifically geared at Martian aurora observation. An aurora is a natural light display in a planet's atmosphere, caused by charged particles colliding with atmospheric gases and emitting light. If you've heard of the Aurora Borealis (aka the "Northern Lights"), here on earth, that's probably the most famous example. The authors' system would allow users to identify auroras, classify their types, and analyze their properties. Let’s see what they built and how they built it. At a high level, their system is a data-driven image analysis pipeline. Rather than manually...
Watch the Trailer
Release Date: November 27th, 2024
In this paper they build a runtime anomaly detection framework that gets deployed alongside an embedded application and operates as a real-time monitoring system. The application monitors the target application's memory access patterns, and uses a combination of Discrete Cosine Transform (DCT) and Hidden Markov Models (HMMs) to detect deviations from normal behavior. Let's dig into the details and see how this works. Memory reference sequences are detailed logs of memory access events produced by an application during its runtime. Each sequence is essentially a timeline of memory addresses that the program reads from or writes to, in the exact order they occur. These events include operations such as accessing variables, loading data into cache, or interacting with external storage. By capturing...
Watch the Trailer
Release Date: November 26th, 2024
A complex number is a number of the form a + bi, where a and b are real numbers, and i is the square root of -1. Here, a is the real part, and b (when paired with i) is the imaginary part. Together, they form a point in a two-dimensional space called the complex plane, where the x-axis represents the real part and the y-axis represents the imaginary part. This allows operations like addition, multiplication, and rotation to be visualized geometrically. In many signal processing applications, such as telecommunications, radar, and image processing, the representation of data as complex numbers is not just convenient but necessary. Complex numbers are used to...
Watch the Trailer
Release Date: November 25th, 2024
This is a system design paper in which they envision a system that can help Podman monitor, intercept and analyze the syscalls happening from within semi-trusted or untrusted containers. This is about telemetry and observability, nothing more. The authors make the case that there may be instances where you need to run a container and give it access to the ability to make syscalls (which it wouldn’t have by default), but then need to monitor what it’s doing carefully. This could be a piece of vendor software, or a piece of software you wrote that incorporates a 3rd party library you haven’t fully vetted yet. It might need access to syscalls for some reason we don’t...
Watch the Trailer
Release Date: November 24th, 2024
The real contribution in this paper is a framework called Modified Regularization Long Short-Term Memory (MR-LSTM). There’s a lot going on in that name, so we’re going to need to build up a definition slowly. A Long Short-Term Memory (LSTM) network is a specialized type of recurrent neural network (RNN) designed to process sequential data. Unlike traditional neural networks, LSTMs excel at retaining long-term dependencies in data sequences, thanks to their unique gating mechanisms. These gates—input, forget, and output—control the flow of information, selectively remembering or discarding details to maintain context over extended timeframes. This makes LSTMs particularly adept at...
Watch the Trailer
Release Date: November 23rd, 2024
A frequency distribution revealed that 46% of studios adopted competitive analysis as a UX technique. Cluster analysis grouped studios based on shared characteristics, identifying patterns such as the fact that small studios often cited financial limitations and lack of knowledge as barriers to UX. Temporal analysis showed a shift over time, with newer studios (founded post-2021) demonstrating greater reliance on MVPs and prototypes for validation while older studios (pre-2015) more frequently employed structured UX practices and dedicated teams....
Watch the Trailer
Release Date: November 22nd, 2024
In today’s paper, the authors propose a new concept: a new variant of ResNet that overcomes two key issues with that methodology. They’re calling their concept a “Quotient Network”. Where ResNet learns in part by calculating differences between old and new features, the Quotient Network learns by calculating the quotients between them instead. So rather than subtraction they’re doing division. Now, that probably makes no intuitive sense to you right now, and that’s ok. We’re about to change that....
Watch the Trailer
Release Date: November 21st, 2024
The pipeline begins with preprocessing. Raw audio recordings often contain noise—unwanted artifacts that can distort the data. So the first step is to employ a Chebyshev filter, known for its sharp roll-off characteristics, to minimize high-frequency noise while retaining the integrity of the signal. For instance, this might clean up background sounds like static or muffled speech, ensuring the emotional cues within the user's voice remain prominent. Once the signal is cleaned, it is segmented using a process called framing. Audio data is continuous, but machine learning models require discrete ...
Watch the Trailer
Release Date: November 20th, 2024
Before we talk about today’s paper, we need to pause, jump in our time machines, and go back to talk about a paper from 26 years ago. In 1998, Yann Lecun, a professor at NYU, published a seminal paper on computer vision: Gradient-Based Learning Applied to Document Recognition. In that paper, the authors outlined the application of convolutional neural networks (CNNs) for recognizing handwritten characters. But more importantly for our purposes today, that paper was also the vehicle through which they first published the MNIST dataset. If you’ve worked...
Watch the Trailer
Release Date: November 19th, 2024
If you’re working as a network engineer for a high traffic application, like a website or an API your mandate is likely pretty broad. You need to not only accommodate, serve, log and analyze today’s traffic, but you need to preemptively build the systems that will allow you to continue doing those things tomorrow. If the site scales up, if the traffic peaks, if there’s a flash-sale or a news article that changes the traffic patterns you need to be on top of it…before it happens. And you need to do all of this without a crystal ball. How?...
Request this Episode
Release Date: November 18th, 2024
The system employs a one-stage object detection model, YOLOx, as its foundation. Unlike two-stage detectors that rely on region proposal networks (RPNs) to generate candidate object regions—an inherently computationally intensive process—YOLOx bypasses this stage by directly predicting bounding boxes and classifications in a single pass through the network. This design allows for faster inference speeds, a critical feature for real-time applications. The YOLOx model serves as a lightweight yet robust baseline, leveraging CSPDarkNet as its backbone for efficient feature extraction. This backbone strikes a balance between computation and feature representation, ensuring sufficient granularity to detect objects under varying conditions of motion and occlusion....
Watch the Trailer
Release Date: November 17th, 2024
They started with YOLOv8, a single-stage object detection model designed for efficient and fast object recognition. Its architecture consists of three primary components: The backbone, the neck, and the detection head. The backbone serves as the feature extractor, processing raw image data to produce rich feature maps that emphasize important aspects of the input, such as edges, shapes, and textures. The neck is responsible for feature fusion, taking information from different levels of the backbone to combine spatial and semantic details into more refined representations. Lastly, the detection head predicts bounding boxes and class labels based on these fused features, outputting the final detection results....
Watch the Trailer
Release Date: November 16th, 2024
China is producing nearly 40 million tons of apples a year, and there’s simply not enough people to pick them. So, Chinese producers are increasingly turning to apple-picking robots to do the job. These can take several different forms, but the one we’re looking at today is a single arm on a platform with an effector attached that can grab apples, twist them off the branch, and plunk them into a basket....
Watch the Trailer
Release Date: November 15th, 2024
This paper is about two very different, but related things. On one hand, yes, it is about using Machine Learning to predict or diagnose the existence of Polycystic Ovary Syndrome. But on the other hand, it’s really about explainable A.I. You see, researchers have been developing ML models to predict or flag PCOS in patients for years. It’s such a popular area of research that there are openly available Kaggle datasets on the subject that allow anyone to easily train a model to do that. It’s been done, and it’s being done. That’s really not the point here....
Watch the Trailer
Release Date: November 14th, 2024
In 1916 Albert Einstein predicted the existence of Gravitational Waves: disturbances or ripples in spacetime. Transient displacements in a gravitational field. Gravitational Waves (GWs), would be caused, he said, by massive, accelerated objects like colliding black holes or neutron stars. But, he also predicted that GWs would be extremely difficult to measure, requiring incredibly sensitive instruments....
Watch the Trailer
Release Date: November 13th, 2024
In 1998, NASA launched the Mars Climate Orbiter. The spacecraft was designed to orbit Mars and relay vital atmospheric data back to Earth. For months, the orbiter traveled through space, with mission-control monitoring its progress and making minor adjustments as needed. Anticipation built as the team prepared for the critical moment when the orbiter would enter Mars' orbit. On the scheduled day, engineers gathered together, awaiting confirmation that the spacecraft had successfully positioned itself around Mars. Instead, they were met with an unsettling silence. Attempts to contact the orbiter failed, and it became clear that the Mars Climate Orbiter was lost, its mission ending abruptly as it vanished into the Martian atmosphere....
Watch the Trailer
Release Date: November 12th, 2024
Caching is, in a word, complex. And in today's paper, it gets significantly more complex. Some would say it becomes ridiculously complex, but I’ll leave that for you to decide. This paper takes the traditional concept of a "cache-hit" or "cache-miss" and adds a third option: a "similarity-hit". A similarity-hit means that the system found cached content that is (in some way) close to what the user is requesting, but not exactly it. Instead, it’s something similar-enough that it might still satisfy the user, and it can be retrieved from the cache rather than the original source....
Watch the Trailer
Release Date: November 11th, 2024
Today, November 11th 2024, marks the first day of COP29, the 29th meeting of the Conference of the Parties. COP was created by the UNFCCC, the United Nations Framework Convention on Climate Change, to bring countries together to negotiate and set international climate action goals. COP29 is happening in Baku, Azerbaijan, and will last nearly two weeks. During this time, member countries will set their climate change policies for...
Watch the Trailer
Release Date: November 10th, 2024
In the early days of the 20th century, AT&T had a problem. They wanted to build the first transcontinental phone line, connecting San Francisco and New York City. But, their engineers told them it was practically impossible. Voices in a phone line couldn’t travel anywhere near that far. Why? Signal attenuation. As signals flowed through the phone lines over distance, they got weaker and weaker. Even if they had the physical wires connecting the two...
Request this Episode
Release Date: November 9th, 2024
Imagine that you're a bank, and you've got a fraud problem. People are signing up for your credit-cards with stolen identities, racking up charges, then disappearing. And it's costing you a fortune. It turns out, the fraudsters are doing this all over town, all over the state, all over the country. It's not just a problem for your bank, it's a problem for all the others as well. So one day you and the other bankers decide to get ahead of it....
Watch the Trailer
Release Date: November 8th, 2024
Let’s say you’re a software engineer at a major video-streaming company. This could be Netflix, Max, YouTube, or even Twitch or Kick. Either way, your company’s core competency is video: ingesting it, processing it, encoding and decoding it, hosting it, and streaming it out to consumers. And you’re good at it. But a new feature request just came down from on-high, and it’s, well, as we like to say: "non-trivial." What feature? Search....
Watch the Trailer
Release Date: November 7th, 2024
If you’re a lover of fine art, I have a little bit of a spoiler for you today. Some of the most iconic pieces are not actually fully original anymore. If you go see the Mona Lisa, The Last Supper, Guernica, Starry Night, or even the ceiling of the Sistine Chapel, you’re not seeing that art as the painter left it however many decades or centuries ago. You’re seeing a version of the art that curators and historians have attempted to revive and restore to reflect what it looked like back when it was first painted....
Watch the Trailer
Release Date: November 6th, 2024
If you ever rent an Airbnb out in the desert somewhere, like Joshua Tree, for example, your host is probably going to spend quite a bit of time talking to you about your shoes. Where to put them, how to put them there, where not to put them, what to do before you put them on, and what to do if you ignored the first instructions and accidentally left them somewhere you weren’t supposed to leave them....
Watch the Trailer
Release Date: November 5th, 2024
What makes good code? If you have two snippets that solve the same problem without throwing errors, what makes one of those snippets better than the other? More advanced than the other? More maintainable than the other?...
Watch the Trailer
Release Date: November 4th, 2024
Think back for a moment to the first time you ever drove a car. It was you, it was whoever was teaching you, the steering wheel, gas, brake, clutch, and the shifter. If you’re a bit younger, then maybe no clutch. But as far as advanced technology, that was probably it. You may have had a radio, or some way to play music, but if you're like me you didn't have a phone, GPS, lidar, cameras, motion sensors, traffic data, bluetooth, or anything else fancy....
Watch the Trailer
Release Date: November 3rd, 2024
Late on the night of March 12, 1928, the St. Francis Dam, a crucial piece of Los Angeles’s water infrastructure, failed without warning. Built just a few years earlier, the dam had been a symbol of progress, an ambitious project to secure water for a growing city. But at 11:57 p.m., a crack turned into a rupture, and within seconds, the dam broke apart, releasing billions of gallons of water into the San Francisquito Canyon. A wall of water, over 100 feet high, surged downstream, taking homes, farms, and hundreds of lives with it. By the time it reached the Pacific Ocean, over 50 miles away, the flood had left a trail of destruction across the valley....
Watch the Trailer
Release Date: November 2nd, 2024
A couple of months ago, a video went viral on TikTok of a patient named Jessica explaining the lengths she has to go to just to get her various medical providers on the same page. Jessica carries a huge binder to her medical appointments, and from the sounds of it, that binder contains virtually every x-ray, procedure, medication, lab result, and exam they've ever had. It includes everything that any doctor who has seen them has ever documented, prescribed, ordered, or said about Jessica. Some people might watch that video and think: Why on earth is Jessica doing this? Why is this necessary? But for the millions of people who have watched and rewatched this video, for the numerous people in the comments who have asked Jessica how to structure their own patient binder, what information to include, and whether to use a cover page, Jessica really strikes a chord....
Request this Episode
Release Date: November 1st, 2024
If you go to Google Trends or the Keyword Planner and look up the worldwide search traffic for the word “Kafka,” you might be forgiven for assuming that Metamorphosis must be on the bestseller list. Every month, there are almost 700,000 searches for that term. What gives? Did every high schooler in the world get assigned the same book report? No, searches for Kafka have been rising for a decade and started spiking a year ago not because of Franz Kafka but because of Apache Kafka, the open-source distributed event streaming platform....
Watch the Trailer
Release Date: October 31st, 2024
This paper has a whole lot going on, and from the title, it might not be entirely clear what it’s focusing on. At the highest level, this paper is about creating a better system for Software Fault Prediction (SFP). That is: shoring up the quality of the software you deliver by using machine learning model that can look at your codebase and predict where issues will arise. SFP isn’t by itself a holistic Quality Assurance process, but it can be a meaningful part of a larger QA/QC regime. So everything we’re going to talk about in this paper does in some way tie back to that: creating better SFP tools so that developers can ship better code....
Watch the Trailer
Release Date: October 30th, 2024
Just outside of Geneva, right near the border between Switzerland and France is the charming little municipality of Meyrin. Population about 26,000. They’ve got some apartment buildings, some low-rise office parks, a few hotels, some restaurants, cafes, a train station, and oh yeah… they’ve got the world’s largest and highest-energy particle accelerator: the LHC. CERN’s Large Hadron Collider. If you’re driving around, you won’t notice the LHC at all because it’s actually buried about 100 meters underground. This 17-mile track has been home to some of the most momentous experiments in modern physics. If you remember the race to identify the Higgs boson subatomic particle a few years ago...that was here. They did that. And that’s just one of the types of experiments that the LHC does....
Watch the Trailer
Release Date: October 29th, 2024
Imagine it’s 2003, and your friend brings a thumb-drive to school. On it: a pirated version of Outkast’s new album, Speakerboxxx/The Love Below. He was up all night on Limewire finding and downloading a decent copy. But he did it. This copy is solid, he listened to the whole thing on his PC before he copied it onto the drive. You borrow the thumb-drive and take it home, and plug it into your Mac. Error. You try again. Error again. Your Mac can’t read the drive. Something about NTFS. The computer suggests you reformat the drive, a process that will erase all the files. You stare at the screen, no Roses for you today....
Watch the Trailer
Release Date: October 28th, 2024
Let’s say I am a despot at the head of an oppressive totalitarian government. You are one of the millions of people who live in the country under my control. One way I can retain my position and stay in power is through mass surveillance, keeping an eye on what everyone is saying, what they’re doing, where they’re going, and who they’re associating with. A big part of that is monitoring what they’re doing online: what sites they’re visiting, how long they’re staying on those sites, and how much information they’re sending to and receiving from those sites. If I can figure out every website that everyone is visiting and every app they’re using, I can use that data to crack down on journalists, dissidents, activists—you name it. Even if I can’t see the information they’re sending to or receiving from these sites, just knowing the sites they’re connecting to is enough for me to act....
Watch the Trailer
Release Date: October 27th, 2024
If you’ve ever tried to train a deep neural network, you’ve probably spent a lot of time waiting. Waiting for training, waiting for boosting, waiting for validation runs, waiting waiting waiting. Why is that? Well, remember in school when you learned matrix multiplication, and you had to do it by hand? Remember how long that took? How tedious it is? Well yeah, your computer feels the same way. The crux of the problem here are MACs: Multiply-Accumulate Operations, in which matrix multiplication is an integral part. MACs are what the machine is doing, and what’s taking so long....
Watch the Trailer
Release Date: October 26th, 2024
If you’ve been listening to Journal Club for a while, you’ve probably noticed that a lot of the Machine-Learning, Deep Learning and Computer Vision episodes have a similar story arc:The researchers have a problem. They collect data. They clean and normalize that data. They extract and select the features they care about. They pick a few algorithms to test against each other. They run the training. They choose a few metrics to use to validate the models and benchmark them against each other. Sometimes they also stack models together or use some kind of ensemble learning. That storyline exists in Journal Club so frequently because that is the most common flow for how this research happens. But what our storylines don’t necessarily capture, is that many of these steps are highly manual, tedious, and repetitive. I might, for example, casually mention that a researcher trained models on KNN, Random Forest and SVM, but that doesn’t mean that those three training cycles looked anything like each other. They might have been in totally separate programs or notebooks, with different libraries, happening at different times on different machines, with a number of discrete manual idiosyncratic steps for each one. Getting different models trained with different algorithms for a given set of data isn’t generally a push-button thing....
Watch the Trailer
Release Date: October 25th, 2024
When I go on vacation, I like to do: nothing. I want to fly to wherever I’m going, sleep all day, chill all night, lay around the room, lay around the pool, order room service, be a sloth. For me, the idea of having to wake up early, put on actual clothes, and go do something sounds completely insane. Why anyone would jam-pack their vacation time with an itinerary full of stuff they have to do is beyond me. That just sounds like work....
Watch the Trailer
Release Date: October 24th, 2024
We all use search engines every day, for a variety of tasks. And our use of them can be broadly categorized into two types: standard searches and exploratory searches. In a standard search, a user looks up information when they already have some understanding of the subject. Even if they don't know the specific answer, they know enough to craft a coherent query that leads to relevant results. For example, searching for "what is the capital of France?" is straightforward because the user knows the general structure of the answer (a single city name) and can phrase the question clearly. This is the type of search traditional engines like Google excel at—they deliver concise, accurate results based on well-defined, unambiguous queries....
Watch the Trailer
Release Date: October 23rd, 2024
Let’s say you’re in charge of network security at a decent sized eCommerce site. Your company’s big enough that they don’t use AWS or GCP or any other cloud provider, they just run their own machines, maybe in a datacenter or colo-center, or maybe even in a server closet right next to your desk. Either way, that network is a real physical thing that your company relies on, and everyone’s counting on you to protect it. Everything’s going fine for a while. But then, a week before your biggest sale of the year—a sale that accounts for a huge percentage of your company’s annual profits-–-a week before that, it happens....
Watch the Trailer
Release Date: October 22nd, 2024
Toe-to-toe, a little honeybee doesn’t stand a chance against a hornet. Most hornets are 2-3 times the size of a bee, plus they have a thick exoskeleton like a lobster. And they’re mean. When a hornet enters a honeybee’s nest, it’s not there to make friends. It’s there to kill bees and steal food. But honeybees have a secret weapon: a technique called "balling". The bees latch onto the intruder one by one, eventually surrounding it in a blanket of bees. As they hold on, they vibrate their flight muscles, generating heat. Lots of heat. The hornet gets hotter, and hotter, and hotter, and then it dies. Victory: bees....
Watch the Trailer
Release Date: October 21st, 2024
Let’s say you and I are Oompa Loompas in a chocolate factory. We have various jobs, and we do all the meaningful work, while the crazy guy in the hat gets all the credit. But that’s fine, it’s what we signed up for. Your job is to take caramel squares and dip them in chocolate. Then you hand them to me. I sprinkle a little salt on top and carefully wrap each chocolate-covered caramel in a cellophane wrapper, twist the ends, and then place that carefully in a box lined with tissue paper. You dip, I wrap....
Watch the Trailer
Release Date: October 20th, 2024
In September 2023, Neuralink received approval to begin human clinical trials for a neural implant called the N1. Their study, nicknamed “Prime,” builds on years of animal trials conducted on pigs and monkeys. These studies were so revolutionary and buzzworthy that even the participants, the animals themselves, became internet celebrities. Millions of people watched as Pager, the monkey, used his Neuralink implant to play ping pong with his mind....
Watch the Trailer
Release Date: October 19th, 2024
In late October 1997, Atlanta Georgia played host to a 4-day technology conference: the 1997 International Conference on Network Protocols. At that event, a paper was presented by a 3-person team from AT&T Research Laboratories, Lucent Technologies Bell Labs, and Fujitsu Laboratory of America. AT&T, or “Ma Bell,” had been broken up a decade before in antitrust proceedings, but these ostensibly now-independent entities still collaborated on research. The paper they presented had all the hallmarks of Bell Labs innovation, namely being decades ahead of its time. The paper was called “A cooperative packet recovery protocol for multicast video,” and in it, they outlined a system in which packet loss between a sender and a receiver could be mitigated by a 3rd party server that steps in to replace packets as they’re lost....
Watch the Trailer
Release Date: October 18th, 2024
In the 2017 mayoral race in Jackson Mississippi, an unusual thing happened. The democratic incumbent, Mayor Tony Yarber, was challenged by members of his own party. Eight of them. They attacked him on his record, on his policies, and on his plans. But most of all, they attacked him on potholes. You see, potholes are endemic in Jackson. About 200 miles north of New Orleans, Jackson is often hot and humid. The moisture penetrates the road surface, seeping into small cracks, and weakening the foundation. Over time, the soil is less able to support the road surface, and as traffic passes over it, deformities begin to form. First rutting, then larger dips and cracks. Meanwhile, scorching midday temperatures soften the road, then colder nights cause it to contract again, and the cracks and dips get worse. Water fills the deformities and the vicious cycle continues. In Jackson that meant pothole after pothole after pothole. The residents had seen enough, and were ready for someone to step in and fix the problem....
Watch the Trailer
Release Date: October 17th, 2024
In 2020, Apple revealed the first release in their “M” series of computers, their foray into designing and building their own chips. No longer would they depend on Intel's processors to power their machines; the future of Apple would be vertically-integrated chip development. And it all started with The M1. While most of the headlines of the time were about the sheer processing power of the computer, many articles buried the real lead. The incredible part of the announcement was less about Apple bringing chip-manufacturing in-house, and more about the fact that they had switched from x86 to ARM. This wasn’t just a coup d’etat for ARM, it was a win for RISC: The Reduced Instruction Set Computer architecture....
Watch the Trailer
Release Date: October 16th, 2024
The past decade has seen nothing short of a revolution in privacy protection. From GDPR in Europe to CCPA in California, to third-party cookie restrictions built into iOS, and stronger browser defaults around HTTPS. Compared to the wild-west days of the early and mid-2000s, the internet consumer today has far more protections and avenues for recourse than ever before. The operators of web applications, mobile apps, APIs, newsletters, and other internet services are governed by strict compliance regimes that dictate what they can and can’t do with your personal information. Importantly, these regimes also specify how they must care for, handle, and protect your personal data—securing it from bad actors, encrypting it, and ensuring its safety....
Watch the Trailer
Release Date: October 15th, 2024
In the fall of 2021, after nearly six months in the halls of Congress, the Infrastructure Investment and Jobs Act (IIJA) was finally signed into law. This package, also known as the Bipartisan Infrastructure Law (BIL), allocated $1.2 trillion of funding, with $65 billion of that going specifically towards modernizing America’s power grid. Of that $65 billion, around $13 billion (20%) was set aside just specifically for power grid security. That’s thirteen Billion dollars to secure a system that, if you’re like me, you might’ve thought was already secure. Turns out, we were wrong. Not only is the grid vulnerable to a wide range of attacks from large state actors, but as the grid modernizes into a "smart grid," the attack surface actually grows...
Watch the Trailer
Release Date: October 14th, 2024
Okay, so this almost never happens. 99% of the time, on Journal Club the papers we are covering are advancements in computer science. That’s our bread and butter. But once in a blue moon, an article comes out that is the complete opposite. An article that says, in effect: "Hey everyone: when it comes to this one particular technology, we have all gotten out over our skis. We as an industry are making claims that a technology is capable of something it’s not actually capable of." These kinds of articles are rare, but I think they contribute just as much to our understanding of the state of the industry, as the other articles do. So it’s in that spirit that I bring you today’s paper:...
Watch the Trailer
Release Date: October 13th, 2024
Let’s pretend that your name is Janet and you are the CEO of Janet’s Bicycle Company. You sell bicycles to distributors, and you make those bicycles in a factory. The raw materials and third-party components come into the shipping-dock, and you’ve got a production/assembly line where you transform that pile of metal and rubber into a finished bike. From the start of the industrial revolution all the way up to the 1960s, the individual tasks involved in a production-line like yours were manually done by a human being. That person would likely have been using tools, and those tools might have provided leverage, but at its core industrial production was still a human-centered manual task. Bicycles, toasters, cars, shoes, and tape-measures, these were all made by a real person on an assembly line....
Watch the Trailer
Release Date: October 12th, 2024
In May 2021, the White House issued Executive Order 14028, the "Executive Order on Improving the Nation’s Cybersecurity". This order, among other things, directed NIST (the National Institute of Standards and Technology) to develop best-practices around software supply-chain security. Before we go on, I want to disambiguate that phrase. We are not talking about supply-chain software security. That is the security of the software that manages a business' supply chain. No. We’re talking about software supply-chain security: the security of the supply chain of the software we build....
Watch the Trailer
Release Date: October 11th, 2024
Sweet Mama’s Tomato Mix is a tomato-paste produced by the Weddi Africa Tomato Processing and Agro Farm. Weddi is a vertically integrated wholesaler: it operates the farm, it runs the processing plant, and it owns the distribution centers. It delivers five products (pastes in various-sized containers) directly to retailers in its own fleet of trucks, driven by its own drivers. The company operates in Ghana, in or around the city of Kumasi, a metropolitan area that is home to nearly 4 million residents. Weddi's main factory is about a hundred miles outside of the city, but its distribution hubs are closer to the city limits....
Watch the Trailer
Release Date: October 10th, 2024
A few days ago it was a friend's birthday. So at some point during the day I pulled out my phone, went to my message threads, found that person, and typed “Happy”. But as I was typing B-I-R-T-H-D-, the auto-suggest on my phone popped up with an emoji of a birthday cake 🎂. So I clicked on the suggestion, and it replaced the word I was typing with the cake. So what I was left with read as “Happy [Cake]”. I thought to myself, "Why would it think I wanted that?" Why would it think I was replacing the word Birthday with an emoji of a cake? Clearly I wanted to complete the word “Birthday” and then put a cake afterwards. Isn’t that how everyone uses emojis? As decorative accents after the text?...
Watch the Trailer
Release Date: October 9th, 2024
Here’s an uncomfortable thought about your ISP. (Your Internet Service Provider). When it comes to privacy, your interests and their interests are not aligned. Why? Because from their perspective, they need to provide you with a certain quality of service. In order to do that, they need to perform load balancing, routing, traffic prediction and forecasting. They need to identify malicious traffic flows and DDOS attacks. And if you live in a jurisdiction which doesn’t practice net neutrality, then part of their service may include application-specific metering. All of these things are predicated on their ability to classify the traffic that is passing through their network....
Watch the Trailer
Release Date: October 8th, 2024
Imagine for a second that you are an IoT device. You’re a smart watch, or a smart speaker, or a smart lamp, or a smart fridge. Either way you’re smart, you’ve got a processor onboard, and you can handle a lot of the computational workload yourself. But for some tasks, there are just so many calculations to do (so quickly) that your onboard chip isn’t up to the job. For those kinds of workloads, you offload the processing to a remote server, let that machine do the heavy lifting, and then return the result. But there’s a problem. When it comes to choosing the type of server that you send your workload to, you are spoiled for choice. You’ve got three main classes of options: Cloud Servers, Edge Servers, and Fog Nodes....
Watch the Trailer
Release Date: October 7th, 2024
In 1948, less than 90 years after Darwin published "On the Origin of Species", Alan Turing applied the concept of natural selection to Computer Science. In an unpublished essay titled “Intelligent Machinery,” he wrote: “There is the genetical or evolutionary search by which a combination of genes is looked for, the criterion being the survival value.” This line appears to be the first mention by any person, anywhere, of what is now referred to as an evolutionary algorithm. And now, 76 years later, this niche concept which has existed on the fringes of computer science for decades, is poised to finally have its time in the spotlight. Specifically, one technique within this field: Genetic Programming is coming to the forefront. Why? Because the same hardware advances, the same distributed systems and cloud computing advances that have enabled the proliferation of LLMs can also enable the widespread adoption of Genetic Programming....
Watch the Trailer
Release Date: October 6th, 2024
About a month ago, on September 7th, Héctor Garzó rounded the final turn at San Marino, crossed the finish line, and became the 2024 MotoE World Champion. He wasn’t riding any normal motorcycle. He was riding a fully-electric Ducati superbike, with a top speed of about 171 miles per hour. In fact, everyone in that race was riding a Ducati, because the storied Italian manufacturer has signed a deal to be the exclusive provider of electric motorcycles for the entire field through 2026. You see, Ducati is going long on electric motorcycles, and doing everything they can to push the technology (and the acceptance of it) to higher and higher levels. So far, they haven’t seen much success. The MotoE race where Garzo was crowned champion was streamed live on Youtube. As of this writing, the video only has about 11,000 views. Total. For a world championship....
Watch the Trailer
Release Date: October 5th, 2024
In the immortal words of Frankenstein’s monster: FIRE BAD! And unfortunately, we often don’t spot fires until they’re really bad. We’ll miss the small brush fires and only take notice when they’re large and out of control. This happens for a few reasons: Fires can start anywhere, even away from the cities. In the middle of open land, or in a forest where nobody will notice. If someone does notice, there can be a “bystander-effect”: everyone thinks that someone else must have called 911 to report the fire, or that someone else must already be handling it. So nobody ends up calling at all, and the fire grows and grows. Where I live in California, this is a big problem. Large parts of the state are sparsely populated open areas with lots of fuel for wildfires (dry grass and the like). So every summer a sad ritual repeats itself here. Loose cigarette butts, mismanaged campfires, downed power lines, or any of several other causes can be the trigger. And once it starts, the fire goes unspotted for a while, It grows, it spreads and eventually causes real damage. These fires cost California billions of dollars, every single year....
Watch the Trailer
Release Date: October 4th, 2024
NPM is a fantastic package manager in a number of ways, but it's really bad at search. To be fair, every other package manager is bad at it too: PIP, Crates, Maven, Homebrew, RPM, etc. They all struggle to give meaningful search-results for a query. When you have a problem that you need a library to solve, your chances of finding a relevant package might hinge on your ability to guess what that package might be named. The search-bar in NPM returns such irrelevant results, that you'll likely spend your day jumping from search engines, to Github, to blog posts, to youtube videos, to forum discussions. You'll go back and forth to NPM over and over again, trying out different packages to see what fits. You'll spend time reading their docs, checking their Github issues, checking if they're stable, well maintained and recently updated. It’s a pain, to say the least. I’ve always thought that non-programmers would be shocked to see how much of a Software Engineers day is spent trying to find and use new packages. Out of exasperation, many developers just choose the most popular package that seems like it might possibly do the job. Not the package best suited to the problem, not the package with highest test coverage, not the package with the best documentation, or the lowest number of open issues, just the one with the most impressive download graph....
Watch the Trailer
Release Date: October 3rd, 2024
Shenzhen is a bustling metropolis. 17.5 million people, millions of homes, millions of cars, millions of workplaces. But it wasn’t always this way; in fact it wasn’t even this way recently. Thirty years ago Shenzhen was a 10th of its current size, forty years ago it was a quarter of that. The growth experienced in this region over the last half-century has been incredible, and the city is still growing so quickly that local municipalities are having a hard time even tracking the growth. This makes the day-to-day practice of governing somewhat difficult. Common tasks like land-use classification, population density estimation, and even housing-price tracking have become non-trivial. The census is only conducted in China once every 10 years, so how are they supposed to keep track of the population in the meantime?...
Watch the Trailer
Release Date: October 2nd, 2024
First, we need to step back in time for a moment. In 2017, Vaswani et al published a seminal paper that would send shockwaves through the A.I. community: “Attention Is All You Need”. In it, the authors described a novel concept: the Transformer. Transformers (and their building-blocks called “transformer-blocks”) were a new type of a Neural-Network architecture based on the concept of self-attention. Self attention allowed for two key benefits: 1. The ability for tokens to maintain relationships with long-range dependencies. 2. The ability to parallelize both your training and inference. In other words, transformers made Large Language Models possible, set the stage for what would become the Foundation Models, ChatGPT, and arguably the whole AI boom of the last few years. I’m telling you this because the paper we’re about to dive-into builds on top of the concept of a transformer, utilizing an even newer concept: a Conformer. You see, after a few years working with transformers, researchers started to realize that they had a few limitations. Namely, transformers were great at determining the long-range relationships between tokens, but they struggled with short-range relationships or “local dependencies”. So the conformer was born, which utilizes a transformer but adds on convolutions for local patterns. Thus the name “con” from convolution and “former” from transformer. Conformers are designed to be great at both the big picture and the tiny details....
Request this Episode
Release Date: October 1st, 2024
In 2014 Masayoshi Son, the founder of SoftBank, revealed Pepper: A four foot tall humanoid robot that could talk, answer questions, and perform a variety of pre-programmed actions. Pepper had a tablet sticking out of its chest to display visual information to the person it was communicating with. Over the last decade thousands of these little robots were sold around the world and reprogrammed to become everything from front-desk receptionists, to tour guides, to waiters. Most of these installations were experiments: cute demos and a sneak-peak of what might be possible in the future....
Request this Episode
Release Date: September 30th, 2024
Cotton farmers in Pakistan have a problem: The Silverleaf Whitefly. This little insect is a carrier of the Cotton Leaf Curl Virus (CLCuV), which causes Cotton Leaf Curl Disease (CLCuD). Cotton Leaf Curl is serious. Once infected, the virus causes photosynthate blockage in the veins of a cotton plant. Photosynthates are the compounds produced during photosynthesis, so having a photosynthate blockage is really really bad. The tertiary veins on the leaves turn yellow and thicken, the leaves start to curl, then secondary veins get blocked, which reduces surface area of the leaves. This means less photosynthesis, and the downward spiral continues. Infected plants end up shorter, with curly scaly leaves, and most importantly: the stress on the plant can cause cotton yields to drop 80%....
Request this Episode
Release Date: September 29th, 2024
Remember a year or two ago when “passwordless” authentication for websites was all the rage? News articles from the time proclaimed that passwords were dead and that passkeys (and related technologies) were clearly the future. Well…what happened? From what I can tell, a lot of people gave passkeys a try, hated them, and went right back to using passwords. To be fair, passwordless authentication for mobile lock-screens and mobile payments did become a reality (in the form of biometrics) but passwordless authentication on the web never really caught-on (outside of Oauth). It turns out, if you want to disrupt something as ubiquitous as passwords you need to offer a solution that is waaaay better. Not just a slight improvement, but a full step-function better. And in the minds of many users, passkeys didn't reach that bar....
Watch the Trailer
Release Date: September 28th, 2024
Worldwide, approximately 1/5th of children under the age of five suffer from some sort of physical growth-impairment. That figure has improved virtually every year for decades, but there’s still a lot of work left to do. Growth-impairment has traditionally been viewed strictly as a function of malnutrition, but there is an increasing awareness that there are a multitude of other factors that correlate strongly with growth-impairment. Identifying these risk factors could help clinicians intervene earlier and more effectively. In this research the authors attempt to build a predictive model for growth-impairment: a model that can analyze other household and familial risk factors, and inform medical professionals of any elevated risk....
Watch the Trailer
Release Date: September 27th, 2024
In the world of virtual reality, developers and designers care a lot about something called "embodiment". Embodiment is the feeling or sensation that you’re "in" the character/avatar that you're playing. You are the avatar. You have agency, you have ownership. You have what’s referred to as "visual-motor congruence", and you feel accurately represented by the avatar in the virtual universe. Remember that scene in the Matrix when Neo wants to learn Kung Fu? They plug him into the simulation, and suddenly he’s in a room with Morpheus training and fighting. Well, in that moment Neo knows that what he's experiencing is a simulation, but that's counteracted by the fact that he has “high embodiment” in his avatar. It’s not his actual body, but in a way, it is. He’s not really in that room, but in a way, he is. That's embodiment....
Watch the Trailer
Release Date: September 26th, 2024
You've probably heard of both deforestation and reforestation. But what on earth is afforestation? Afforestation is the process of planting and growing trees where none originally grew (or they haven’t grown in a long time). Afforestation is not replanting a recently-cut forest, it’s creating a forest where there was previously grassland, desert, or even ice, for example. Though it sounds a bit odd, afforestation is a key practice in sustainable timber production. Rather than chopping down existing forests, a company will grow a brand new forest, chop it down, and repeat. But, this technique does have its drawbacks. Commercial afforestation is extremely water-hungry. And as a result, many countries regulate it tightly. In South Africa, where today’s research takes place, you can’t grow a new forest without first being issued a water license. The licenses are granted based on the government's statistical models that show how that proposed crop would affect the overall energy balance, environment and water table. In order to create those models, the government needs to know a key metric for the species of tree you want to grow: Evapotranspiration (ET)....
Watch the Trailer
Release Date: September 25th, 2024
Now, if you’re like me: alarm bells are already ringing. These shootouts happen all the time, and in many cases the experimental design is seriously flawed. In my opinion, this paper is no exception. So rather than take this article at face value (and report their findings as fact), we’re going to take a more critical lens. I am going to present their research and their findings, but I’ll spend much more time than usual on history and context, and then I'll use that to inform counter-arguments to their analysis. This study is imperfect, but it’s not without value. They did unearth some interesting findings, I just don’t think those findings should be taken at face value; so we won’t. The way we'll get the most out of this paper is through a critical, interrogative lens, so let's do that....
Watch the Trailer
Release Date: September 24th, 2024
This year over 90% of the new cars sold in the United States have some kind of driver-assistance technology built in. That means some kind of system that can aide in the steering, braking, or both. This basket of technologies is broadly referred to as autonomous driving, semi-autonomous driving, self-driving or full-self driving etc. There’s a lot of names for it, and a lot of implementations. To keep everyone on the same page, self-driving is classified on the SAE Driving Automation Scale, from Level 0 (minimal or no self-driving) all the way up to Level 5....
Request this Episode
Release Date: September 23rd, 2024
Serverless is great. In a number of ways it's truly a transformative technology. And it's spreading: every year more hosts jump into this space to offer their own Function as a Service (FaaS) platforms: from AWS, GCP and Azure, to the edge offerings available on Cloudflare, Fastly and Akamai, to the open-source options like OpenFaaS. But as with anything, serverless is not without its drawbacks. From the "uncapped costs" problem, to the "how do I run this locally?" problem, there are obviously a few areas where the developer-experience could be improved. And that's to be expected given how nascent the technology still is. But there is one issue that's a deal-breaker for many devs. An issue so off-putting that it actually prevents them from adopting serverless at all. That issue is the "cold start" problem....
Request this Episode
Release Date: September 22nd, 2024
A half-hour drive south west of Rome sits an archaeological site called Ostia Antica. It’s the ruins of an ancient city from two millennia ago. Right in the center of the site is a brick house called “Casa Di Diana” (Diana’s House). It’s believed to originally have been a five-story building used for both residential and commercial activities. Now it’s a series of half-standing walls, chambers and pillars. From an archaeological perspective, Casa Di Diana is important. It’s one of the many treasures left from the Roman empire. A site that the archaeologists who have been studying it would love to share with the world. But, in its current form, it’s simply too fragile to let anyone but trained professionals visit....
Request this Episode
Release Date: September 21st, 2024
Right now, in 2024, writing programs for quantum computers feels like you’re writing code for IBM mainframes in the 1960s. Here’s what it's like: You write your code. You find a vendor with an available quantum computer that has a compiler compatible with your code. You send your code to that vendor in the hopes that they will compile and run your code sometime soon. You wait and you wait and you wait (for minutes or hours or days) You eventually get a readout back of the result of your program and/or your errors. The whole process is like stepping into a time machine!...And not in a good way....
Request this Episode
Release Date: September 20th, 2024
Way back in 2014 Google released an open-source version of Borg, their in-house cluster management system. Borg was immense and highly specific to Google’s systems and servers, but this new open-source version would be different: It would be smaller, portable to different types of systems and hardware, focused narrowly on container orchestration, and generally useful for a number of workloads. They called the new project: Kubernetes. Kubernetes came out at exactly the right time. Adoption of Docker had spiked in the previous years, microservices were all the rage, and companies of all sizes were grappling with a non-trivial task: container orchestration. With no robust turnkey solutions available, developers were either building one-off systems from scratch, or struggling to port their existing CI/CD systems to work with containers. Google was uniquely positioned to step in. They had been early adopters of containers, and had spent a decade learning how to run them at scale. Kubernetes (or k8s as it came to be known) was the distillation of everything they’d learned...just scaled down....
Watch the Trailer
Release Date: September 19th, 2024
For as long as there have been classrooms, there have been students in those classrooms figuring out ways to cheat. And as long as students have been cheating, there have been teachers and administrators trying to catch them in the act. Nothing about this cat-and-mouse game is new. What is new however, is the power of the AI tools that students now have available to them. Foundation Models exposed through interfaces like ChatGPT can produce a decent term-paper for a student in seconds, and this has drastically changed the playing field. So what’s a teacher or professor to do? How can they be expected to identify which papers were actually written by their students, and which were copied-and-pasted from a GenAI tool?...
Watch the Trailer
Release Date: September 18th, 2024
The researchers in this study wanted to compile a database of police-reports, and then group those reports by the type of crime-scene described within each document. Seems simple enough, no? Unfortunately the authors immediately hit a wall: It turns out that police reports contain so much sensitive and private information that they’re rarely made public. So the authors couldn't get their hands on any of them and had to figure out a workaround: Rather than deal directly with police reports, they would use public court documents instead. Their theory was that court documents often have the narrative-areas of the relevant police-reports embedded within them (literally copied and pasted from the police reports right into the court docs). So the researchers could still build a database of police-writing, they'd just have to get that writing from court documents instead of directly from the police reports....
Watch the Trailer
Release Date: September 17th, 2024
If you find yourself working with IoT devices, you’ll probably spend a lot of time thinking about light. No matter how elegantly you program your device, and no matter how cleanly you construct the backend, the interactions between your device and its server will only ever be as quick and responsive as light allows. Light travels at around 300,000 kilometers per second. That's around 186 miles per millisecond. So for tasks that require a round-trip to the server, the ultimate latency and lagginess of your device has nothing to do with your programming abilities; it has to do with how physically far the IoT device is from the nearest data center. Let's say you're building a device with strict latency requirements, and you want a roundtrip baseline of 6 milliseconds. If that’s the case, your server can’t be more than around 560 miles away from the device. Any further and light won’t have enough time to get to the server and come back....
Watch the Trailer
Release Date: September 16th, 2024
One of my earliest memories is of the Loma Prieta earthquake. It hit California on a Tuesday afternoon in 1989. I was three years old and I was at day-care, playing in the sandbox. All I remember is that my universe went from still and quiet to violent and loud very suddenly. In just twenty seconds 63 people died, nearly 4,000 were injured and over 12,000 displaced. A piece of the upper deck of the Bay Bridge collapsed, the Nimitz freeway in Oakland fell down, and local homes and businesses suffered billions of dollars in damages. There was no warning. None. Back then Earthquake Early Warning systems (EEWs) had been invented, but they weren't installed in California until a few years later. So when a big earthquake hit, nobody had a chance to prepare or take cover. Authorities had no time to slow freeway traffic, turn stop-lights red, or close bridges. Now, thankfully that’s all changed. And for good reason! The "big one" (a megathrust earthquake) could happen in California anytime, and is projected to happen sometime in the next few decades. This time, when it hits, we'll at least have a few moments of warning....
Watch the Trailer
Release Date: September 15th, 2024
Picture this: It’s some time in the not too distant future, and you’re out with your friends one night having a few drinks. Maybe a few too many drinks. At some point you decide to leave, jump in your car and start driving down the road. Within a few blocks your car becomes aware that you’re driving drunk and it takes action. What happens next is anyone’s guess: maybe the car takes your control away and goes into full self-driving mode. Maybe it just pulls over and turns itself off. Maybe it calls you a cab, maybe it calls the police. Who knows. The more pertinent question in front of us today is: How on earth did the car know you were driving drunk in the first place?...
Watch the Trailer
Release Date: September 14th, 2024
As lithium-ion batteries age they get worse and worse and eventually reach the end of their useful life. This is a big issue for companies who operate fleets of UAVs (drones). If they replace the batteries in their drones too early, they’re wasting money. If they replace them too late their missions and deliveries get compromised. So they need a way to predict exactly when a drone’s battery is going to reach the end of its useful life. This paper is using AI to do exactly that. Let's take a look....
Watch the Trailer
Release Date: September 13th, 2024
If you’ve been around computer science for a while you’ve undoubtedly heard of the Bin-Packing problem. While it's surely difficult, imagine how much worse it would be if you had to optimize for several variables at the same time. That’s the situation faced by high volume SaaS companies who are trying to allocate Virtual Machines (or containers) onto their bare metal compute. They’re not just trying to minimize resource waste and cost, they need to balance power consumption, failure rates, application responsiveness, uptime, and more. Doing that is NP-hard and incredibly complex. In this paper the authors conceive an elegant solution: A framework called MoVPAAC....
Watch the Trailer
Release Date: September 12th, 2024
In the tech industry we’ve gotten pretty used to disruption. It seems like every few months there’s a new technology that’s "going to change everything". But very few technologies have the potential to be as truly disruptive as autonomous driving. Why? Because in the United States, it’s estimated that around a third of all civilian jobs involve driving in some way. And around 3% of all jobs are full-time driving. So it’s not hyperbole to say that a huge number of American workers can and will have their daily lives and employment significantly impacted by autonomous driving. Now, you may think autonomous driving is a bad thing (there’s an argument for that), or you may think it’s a great thing (there’s an argument for that as well). But what I think we can all agree that technology this disruptive can't and shouldn't be solely in the hands of a few giant companies with closed-sourced systems....
Watch the Trailer
Release Date: September 11th, 2024
Here's a fun fact: your credit score isn't what it used to be. Now, I don't mean that your personal credit score has gone down, I mean that the days when banks relied solely or even mostly on your credit score to determine credit worthiness are gone. Long gone. These days: When a bank needs to determine your credit worthiness, it might look briefly at your score, but then its going to use its own proprietary A.I. to review all of your credit and banking history, as well as all the publicly available data that it can find on you. Then its model will spit out a credit worthiness decision. And lest we forget: credit worthiness determines a lot. It doesn't just determine whether or not you're approved for a loan or credit card, but the interest rate you get, the size of the credit limit, the rewards you qualify for, and more....
Watch the Trailer
Release Date: September 10th, 2024
The field of Psychiatry doesn't always lend itself well to data-science. The field of Computational Psychiatry, on the other hand, thinks we can do better. In this new study, researchers built a proof-of-concept model that should be able to take a new patient (who is admitted for depression, mania or psychosis), review their history, and then project forward the ultimate progression and possible treatment of their condition. Far fetched? Let's dive in and find out....
Watch the Trailer
Release Date: September 9th, 2024
A few companies are planning on putting their datacenters in the stratosphere. Yes, the stratosphere! Their theory is that the sheer amount of free cooling would justify the costs and headache. For a moment let's put aside how they're going to get the machines all the way up there, and then let's put aside how they're going to keep them afloat. There's still one big problem: TCP isn't going to work very well in the stratosphere. Not well at all. That's according to a new study in which the researchers are proposing an alternative new variant of TCP for use exclusively in the stratosphere. They're calling TCP Stratos....
Watch the Trailer
Release Date: September 8th, 2024
In Computer Science we don't talk about cashews very often, but maybe we should. A team of researchers in Vietnam just trained a computer-vision classifier on top of YoloV8 that can grade cashew nuts as they're coming down a conveyor belt. It can perform the grading in milliseconds, with less than a 3% error rate, running entirely on cheap off-the-shelf components. Impressive!...
Watch the TrailerThat's it. You've reached the end of the archive.