Singularity Hub Daily

Latest Episodes:

GE's New Autonomous Electric Pods Have No Steering Wheel, Pedals, or Cab Nov 04, 2021

Show notes

Self-driving cars are taking longer to become a reality than many experts predicted. But that doesn’t mean there isn’t steady progress being made; on the contrary, autonomous driving technology is consistently making incremental advancements in all sorts of vehicles, from cars to trucks to buses. Last week, General Electric (GE) and Swedish freight tech company Einride announced a partnership to launch a fleet of autonomous electric trucks. According to the press release, the fleet will be the first of its kind to operate in the US, with trucks running on GE’s 750-acre Appliance Park campus in Louisville, Kentucky, as well as at GE facilities in Tennessee and Georgia. Einride was founded in 2016, raised $25 million in Series A funding in 2019, and most recently raised $110 million this past May. Though the company makes electric trucks that are driven by humans, it’s primarily known for another vehicle that can’t quite be called a truck; the company’s Pods not only lack drivers, they lack cabs for drivers to sit in. The vehicles are similar to Volvo’s Vera trucks, which also lack cabs, though the Pods are meant for medium distance and capacity, like moving goods from a distribution center to a series of local stores. The Pods will need regulatory approval to operate on public roads, so for now they’ll be limited to driving around within customers’ campuses (a safety driver isn’t required if the vehicles are on private property). According to Reuters, Einride just hired its first remote driver (based in the US) who will be able to take over control of the Pods if they get into a situation where human decision-making is required. Despite being confined to corporate campuses for now, Einride envisions its technology making a measurable difference in terms of emissions, estimating that the GE partnership will save the company 970 tons of carbon dioxide emissions in the first year, with that quantity increasing as more trucks are added. “Between seven and eight percent of global CO2 emissions come from heavy road freight transport,” Robert Falck, CEO and founder of Einride, told TechCrunch. “One of the drivers for starting Einride is that I’m very worried that by optimizing and making road freight transport autonomous, but based on diesel, it’s likely that we will actually increase emissions because it would become that much cheaper to operate.” This would be a textbook example of the Jevons Paradox: when technological progress increases the efficiency with which a resource is used, pushing its cost down—but causing consumption to then rise due to increasing demand, ultimately canceling out any savings of said resource. Electrification, then, is a key aspect of Einride’s strategy, and one of its three core focus points (along with digitization and automation; the company created a digital platform that handles customers’ planning, scheduling, routing, invoices, and billing). “You can actually electrify up to 40 perecent of the US road freight transport system with a competitive business case using existing technology,” Falck said. “It’s more about deploying new way of thinking rather than just improving the hardware.” Einride also has contracts in place with tire maker Bridgestone and oat milk manufacturer Oatly. The company plans to open a US headquarters in New York next year, as well as offices in Austin and San Francisco. Image Credit: Einride

Scientists Mapped Every Large Solar Plant on the Planet Using Satellites and Machine Learning Nov 03, 2021

Show notes

An astonishing 82 percent decrease in the cost of solar photovoltaic (PV) energy since 2010 has given the world a fighting chance to build a zero-emissions energy system which might be less costly than the fossil-fueled system it replaces. The International Energy Agency projects that PV solar generating capacity must grow ten-fold by 2040 if we are to meet the dual tasks of alleviating global poverty and constraining warming to well below 2°C. Critical challenges remain. Solar is “intermittent,” since sunshine varies during the day and across seasons, so energy must be stored for when the sun doesn’t shine. Policy must also be designed to ensure solar energy reaches the furthest corners of the world and places where it is most needed. And there will be inevitable tradeoffs between solar energy and other uses for the same land, including conservation and biodiversity, agriculture and food systems, and community and indigenous uses. Colleagues and I have now published in the journal Nature the first global inventory of large solar energy generating facilities. “Large” in this case refers to facilities that generate at least 10 kilowatts when the sun is at its peak (a typical small residential rooftop installation has a capacity of around 5 kilowatts). We built a machine learning system to detect these facilities in satellite imagery and then deployed the system on over 550 terabytes of imagery using several human lifetimes of computing. We searched almost half of Earth’s land surface area, filtering out remote areas far from human populations. In total we detected 68,661 solar facilities. Using the area of these facilities, and controlling for the uncertainty in our machine learning system, we obtain a global estimate of 423 gigawatts of installed generating capacity at the end of 2018. This is very close to the International Renewable Energy Agency’s (IRENA) estimate of 420 GW for the same period. Tracking the Growth of Solar Energy Our study shows solar PV generating capacity grew by a remarkable 81 percent between 2016 and 2018, the period for which we had timestamped imagery. Growth was led particularly by increases in India (184 percent), Turkey (143 percent), China (120 percent) and Japan (119 percent). Facilities ranged in size from sprawling gigawatt-scale desert installations in Chile, South Africa, India, and north-west China, through to commercial and industrial rooftop installations in California and Germany, rural patchwork installations in North Carolina and England, and urban patchwork installations in South Korea and Japan. The Advantages of Facility-Level Data Country-level aggregates of our dataset are very close to IRENA’s country-level statistics, which are collected from questionnaires, country officials, and industry associations. Compared to other facility-level datasets, we address some critical coverage gaps, particularly in developing countries, where the diffusion of solar PV is critical for expanding electricity access while reducing greenhouse gas emissions. In developed and developing countries alike, our data provides a common benchmark unbiased by reporting from companies or governments. Geospatially-localized data is of critical importance to the energy transition. Grid operators and electricity market participants need to know precisely where solar facilities are in order to know accurately the amount of energy they are generating or will generate. Emerging in-situ or remote systems are able to use location data to predict increased or decreased generation caused by, for example, passing clouds or changes in the weather. This increased predictability allows solar to reach higher proportions of the energy mix. As solar becomes more predictable, grid operators will need to keep fewer fossil fuel power plants in reserve, and fewer penalties for over- or under-generation will mean more marginal projects will be unlocked. Using the back catalogue of satellite imagery, we were able to estimate ins...

These Mice Pups Inherited Immunity From Their Parents—But Not Through DNA Nov 02, 2021

Show notes

The rules of inheritance are supposedly easy. Dad’s DNA mixes with mom’s to generate a new combination. Over time, random mutations will give some individuals better adaptability to the environment. The mutations are selected through generations, and the species becomes stronger. But what if that central dogma is only part of the picture? A new study in Nature Immunology is ruffling feathers in that it re-contextualizes evolution. Mice infected with a non-lethal dose of bacteria, once recovered, can pass on a turbo-boosted immune system to their kids and grandkids—all without changing any DNA sequences. The trick seems to be epigenetic changes—that is, how genes are turned on or off—in their sperm. In other words, compared to millennia of evolution, there’s a faster route for a species to thrive. For any individual, it’s possible to gain survivability and adaptability in a single lifetime, and those changes can be passed on to offspring. “We wanted to test if we could observe the inheritance of some traits to subsequent generations, let’s say independent of natural selection,” said study author Dr. Jorge Dominguez-Andres at Radboud University Nijmegen Centre. “The existence of epigenetic heredity is of paramount biological relevance, but the extent to which it happens in mammals remains largely unknown,” said Drs. Paola de Candia at the IRCCS MultiMedica, Milan, and Giuseppe Matarese at the Treg Cell Lab, Dipartimento di Medicina Molecolare e Biotecnologie Mediche at the Università degli Studi di Napoli in Naples, who were not involved in the study. “Their work is a big conceptual leap.” Evolution on Steroids The paper is controversial because it builds upon Darwin’s original theory of evolution. You know this example: giraffes don’t have long necks because they had to stretch their necks to reach higher leaves. Rather, random mutations in the DNA that codes for long necks was eventually selected, mostly because those giraffes were the ones that survived and procreated. Yet recent studies have thrown a wrench into the long-standing dogma around how species adapt. At their root is epigenetics, a mechanism “above” DNA to regulate how our genes are expressed. It’s helpful to think of DNA as base, low-level code—ASCII in computers. To execute the code, it needs to be translated into a higher language: proteins. Similar to a programming language, it’s possible to silence DNA with additional bits of code. It’s how our cells develop into vastly different organs and body parts—like the heart, kidneys, and brain—even though they have the same DNA. This level of control is dubbed epigenetics, or “above genetics.” One of the most common ways to silence DNA is to add a chemical group to a gene so that, like a wheel lock, the gene gets “stuck” as it’s trying make a protein. This silences the genetic code without damaging the gene itself. These chemical markers are dotted along our genes, and represent a powerful way to control our basic biology—anything from stress to cancer to autoimmune diseases or psychiatric struggles. But unlike DNA, the chemical tags are thought to be completely wiped out in the embryo, resulting in a blank slate for the next generation to start anew. Not so much. A now famous study showed that a famine during the winters of 1944 and 1945 altered the metabolism of kids who, at the time, were growing fetuses. The consequence was that those kids were more susceptible to obesity and diabetes, even though their genes remained unchanged. Similar studies in mice showed that fear and trauma in parents can be passed onto pups—and grandkids—making them more susceptible, whereas some types of drug abuse increased the pups’ resilience against addiction. Long story short? DNA inheritance isn’t the only game in town. Superpowered Immunity The new study plays on a similar idea: that an individual’s experiences in life can change the epigenetic makeup of his or her offspring. Here, the authors focused on trained immunity—the par...

New Optical Switch Is Up to 1,000 Times Faster Than Silicon Transistors Nov 01, 2021

Show notes

As Moore’s Law slows, people are starting to look for alternatives to the silicon chips we’ve long been reliant on. A new optical switch up to 1,000 times faster than normal transistors could one day form the basis of new computers that use light rather than electricity. The attraction of optical computing is obvious. Unlike the electrons that modern computers rely on, photons travel at the speed of light, and a computer that uses them to process information could theoretically be much faster than one that uses electronics. The bulkiness of conventional optical equipment long stymied the idea, but in recent years the field of photonics has rapidly improved our ability to produce miniaturized optical components using many of the same techniques as the semiconductor industry. This has not only led to a revived interest in optical computing, but could also have significant impact for the optical communications systems used to shuttle information around in data centers, supercomputers, and the internet. Now, researchers from IBM and the Skolkovo Institute of Science and Technology in Russia have created an optical switch—a critical component in many photonic devices—that is both incredibly fast and energy-efficient. It consists of a 35-nanometer-wide film made out of an organic semiconductor sandwiched between two mirrors that create a microcavity, which keeps light trapped inside. When a bright “pump” laser is shone onto the device, photons from its beam couple with the material to create a conglomeration of quasiparticles known as a Bose-Einstein condensate, a collection of particles that behaves like a single atom. A second weaker laser can be used to switch the condensate between two levels with different numbers of quasiparticles. The level with more particles represents the “on” state of a transistor, while the one with fewer represents the “off” state. What’s most promising about the new device, described in a paper in Nature, is that it can be switched between its two states a trillion times a second, which is somewhere between 100 and 1,000 times faster than today’s leading commercial transistors. It can also be switched by just a single photon, which means it requires far less energy to drive than a transistor. Other optical switching devices with similar sensitivity have been created before, but they need to be kept at cryogenic temperatures, which severely limits their practicality. In contrast, this new device operates at room temperature. There’s still a very long way to go until the technology appears in general-purpose optical computers, though, study senior author Pavlos Lagoudakis told IEEE Spectrum. “It took 40 years for the first electronic transistor to enter a personal computer,” he said. “It is often misunderstood how long before a discovery in fundamental physics research takes to enter the market.” One of the challenges is that, while the device requires very little energy to switch, it still requires constant input from the pump laser. In a statement, the researchers said they are working with collaborators to develop perovskite supercrystal materials that exhibit superfluorescence to help lower this source of power consumption. But even if it might be some time until your laptop is sporting a chip made out of these switches, Lagoudakis thinks they could find nearer-term applications in optical accelerators that perform specialized operations far faster than conventional chips, or as ultra-sensitive light detectors for the LIDAR scanners used by self-driving cars and drones. Image Credit: Tomislav Jakupec from Pixabay

Animal Evolution: Fossil Discovery Hints First Animals Lived Nearly 900 Million Years Ago Oct 31, 2021

Show notes

Ever wonder how and when animals swanned onto the evolutionary stage? When, where, and why did animals first appear? What were they like? Life has existed for much of Earth’s 4.5-billion-year history, but for most of that time it consisted exclusively of bacteria. Although scientists have been investigating the evidence of biological evolution for over a century, some parts of the fossil record remain maddeningly enigmatic, and finding evidence of Earth’s earliest animals has been particularly challenging. Hidden Evolution Information about evolutionary events hundreds of millions of years ago is mainly gleaned from fossils. Familiar fossils are shells, exoskeletons and bones that organisms make while alive. These so-called “hard parts” first appear in rocks deposited during the Cambrian explosion, slightly less than 540 million years ago. The seemingly sudden appearance of diverse, complex animals, many with hard parts, implies that there was a preceding interval during which early soft-bodied animals with no hard parts evolved from simpler animals. Unfortunately, until now, possible evidence of fossil animals in the interval of “hidden” evolution has been very rare and difficult to understand, leaving the timing and nature of evolutionary events unclear. This conundrum, known as Darwin’s dilemma, remains tantalizing and unresolved 160 years after the publication of On the Origin of Species. Required Oxygen There is indirect evidence regarding how and when animals may have appeared. Animals by definition ingest pre-existing organic matter, and their metabolisms require a certain level of ambient oxygen. It has been assumed that animals could not appear, or at least not diversify, until after a major oxygen increase in the Neoproterozoic Era, sometime between 815 and 540 million years ago, resulting from accumulation of oxygen produced by photosynthesizing cyanobacteria, also known as blue-green algae. It is widely accepted that sponges are the most basic animal in the animal evolutionary tree and therefore probably were first to appear. Yes, sponges are animals: they use oxygen and feed by sucking water containing organic matter through their bodies. The earliest animals were probably sponge-related (the “sponge-first” hypothesis), and may have emerged hundreds of millions of years prior to the Cambrian, as suggested by a genetic method called molecular phylogeny, which analyzes genetic differences. Based on these reasonable assumptions, sponges may have existed as much as 900 million years ago. So, why have we not found fossil evidence of sponges in rocks from those hundreds of millions of intervening years? Part of the answer to this question is that sponges do not have standard hard parts (shells, bones). Although some sponges have an internal skeleton made of microscopic mineralized rods called spicules, no convincing spicules have been found in rocks dating from the interval of hidden early animal evolution. However, some sponge types have a skeleton made of tough protein fibers called spongin, forming a distinctive, microscopic, three-dimensional meshwork, identical to a bath sponge. Work on modern and fossil sponges has shown that these sponges can be preserved in the rock record when their soft tissue is calcified during decay. If the calcified mass hardens around spongin fibers before they too decay, a distinctive microscopic meshwork of complexly branching tubes results appears in the rock. The branching configuration is unlike that of algae, bacteria, or fungi, and is well known from limestones younger than 540 million years. Unusual Fossils I am a geologist and paleobiologist who works on very old limestone. Recently, I described this exact microstructure in 890-million-year-old rocks from northern Canada, proposing that it could be evidence of sponges that are several hundred million years older than the next-youngest uncontested sponge fossil. Although my proposal may initially seem outrageous, it is consiste...

This Spooky, Bizarre Haunted House Was Generated by an AI Oct 29, 2021

Show notes

AI is slowly getting more creative, and as it does it’s raising questions about the nature of creativity itself, who owns works of art made by computers, and whether conscious machines will make art humans can understand. In the spooky spirit of Halloween, one engineer used an AI to produce a very specific, seasonal kind of “art”: a haunted house. It’s not a brick-and-mortar house you can walk through, unfortunately; like so many things these days, it’s virtual, and was created by research scientist and writer Janelle Shane. Shane runs a machine learning humor blog called AI Weirdness where she writes about the “sometimes hilarious, sometimes unsettling ways that machine learning algorithms get things wrong.” For the virtual haunted house, Shane used CLIP, a neural network built by OpenAI, and VQGAN, a neural network architecture that combines convolutional neural networks (which are typically used for images) with transformers (which are typically used for language). CLIP (short for Contrastive Language–Image Pre-training) learns visual concepts from natural language supervision, using images and their descriptions to rate how well a given image matches a phrase. The algorithm uses zero-shot learning, a training methodology that decreases reliance on labeled data and enables the model to eventually recognize objects or images it hasn’t seen before. The phrase Shane focused on for this experiment was “haunted Victorian house,” starting with a photo of a regular Victorian house then letting the AI use its feedback to modify the image with details it associated with the word “haunted.” The results are somewhat ghoulish, though also perplexing. In the first iteration, the home’s wood has turned to stone, the windows are covered in something that could be cobwebs, the cloudy sky has a dramatic tilt to it, and there appears to be fire on the house’s lower level. Shane then upped the ante and instructed the model to create an “extremely haunted” Victorian house. The second iteration looks a little more haunted, but also a little less like a house in general, partly because there appears to be a piece of night sky under the house’s roof near its center. Shane then tried taking the word “haunted” out of the instructions, and things just got more bizarre from there. She wrote in her blog post about the project, “Apparently CLIP has learned that if you want to make things less haunted, add flowers, street lights, and display counters full of snacks.” “All the AI’s changes tend to make the house make less sense,” Shane said. “That’s because it’s easier for it to look at tiny details like mist than the big picture like how a house fits together. In a lot of what AI does, it’s working on the level of surface details rather than deeper meaning.” Shane’s description matches up with where AI stands as a field. Despite impressive progress in fields like protein folding, RNA structure, natural language processing, and more, AI has not yet approached “general intelligence” and is still very much in the “narrow” domain. Researcher Melanie Mitchell argues that common fallacies in the field, like using human language to describe machine intelligence, are hampering its advancement; computers don’t really “learn” or “understand” in the way humans do, and adjusting the language we used to describe AI systems could help do away with some of the misunderstandings around their capabilities. Shane’s haunted house is a clear example of this lack of understanding, and a playful reminder that we should move cautiously in allowing machines to make decisions with real-world impact. Banner Image Credit: Janelle Shane, AI Weirdness

Deciphering the Philosophers' Stone: How Scientists Cracked a 400-Year-Old Alchemical Cipher Oct 28, 2021

Show notes

What secret alchemical knowledge could be so important it required sophisticated encryption? The setting was Amsterdam, 2019. A conference organized by the Society for the History of Alchemy and Chemistry had just concluded at the Embassy of the Free Mind, in a lecture hall opened by historical fiction author Dan Brown. At the conference, Science History Institute postdoctoral researcher Megan Piorko presented a curious manuscript belonging to English alchemists John Dee (1527–1608) and his son Arthur Dee (1579–1651). In the pre-modern world, alchemy was a means to understand nature through ancient secret knowledge and chemical experiment. Within Dee’s alchemical manuscript was a cipher table, followed by encrypted ciphertext under the heading “Hermeticae Philosophiae medulla”—or Marrow of the Hermetic Philosophy. The table would end up being a valuable tool in decrypting the cipher, but could only be interpreted correctly once the hidden “key” was found. It was during post-conference drinks in a dimly lit bar that Megan decided to investigate the mysterious alchemical cipher—with the help of her colleague, University of Graz postdoctoral researcher Sarah Lang. A Recipe for the Elixir of Life Megan and Sarah shared their initial analysis on a history of chemistry blog and presented the historical discovery to cryptology experts from around the world at the 2021 HistoCrypt conference. Based on the rest of the notebook’s contents, they believed the ciphertext contained a recipe for the fabled Philosophers’ Stone—an elixir that supposedly prolongs the owner’s life and grants the ability to produce gold from base metals. The mysterious cipher received much interest, and Sarah and Megan were soon inundated with emails from would-be code-breakers. That’s when Richard Bean entered the picture. Less than a week after the HistoCrypt proceedings went live, Richard contacted Lang and Piorko with exciting news: he’d cracked the code. Megan and Sarah’s initial hypothesis was confirmed; the encrypted ciphertext was indeed an alchemical recipe for the Philosophers’ Stone. Together, the trio began to translate and analyze the 177-word passage. The Alchemist Behind the Cipher But who wrote this alchemical cipher in the first place, and why encrypt it? Alchemical knowledge was shrouded in secrecy, as practitioners believed it could only be understood by true adepts. Encrypting the most valuable trade secret, the Philosophers’ Stone, would have provided an added layer of protection against alchemical fraud and the unenlightened. Alchemists spent their lives searching for this vital substance, with many believing they had the key to successfully unlocking the secret recipe. Arthur Dee was an English alchemist and spent most of his career as royal physician to Tsar Michael I of Russia. He continued to add to the alchemical manuscript after his father’s death—and the cipher appears to be in Arthur’s handwriting. We don’t know the exact date John Dee, Arthur’s father, started writing in this manuscript, or when Arthur added the cipher table and encrypted text he titled “The Marrow of Hermetic Philosophy.” However, we do know Arthur wrote another manuscript in 1634 titled Arca Arcanorum—or Secret of Secrets—where he celebrates his alchemical success with the Philosophers’ Stone, claiming he discovered the true recipe. He decorated Arca Arcanorum with an emblem copied from a medieval alchemical scroll, illustrating the allegorical process of alchemical transmutation necessary for the Philosophers’ Stone. Cracking the Code What clues led to decrypting the mysterious Marrow of the Hermetic Philosophy passage? Adjacent to the encrypted text is a table resembling one used in a traditional style of cipher called a Bellaso/Della Porta cipher—invented in 1553 by Italian cryptologist Giovan Battista Bellaso, and written about in 1563 by Giambattista della Porta. This was the first clue. The Latin title indicated the text itself was also in Latin. This was ...

This Tiny Personal Aircraft Costs Under $100K and Can Take Off From Your Driveway Oct 27, 2021

Show notes

From buses to taxis to ambulances, the number and type of vehicles set to take to the skies in the allegedly near future keeps growing. Now another one is joining their ranks, and it seems to defy classification—it’s not a flying car, nor a drone; the closest to an accurate description may be a flying all-terrain vehicle, or the designation its creators have given it, which is a “personal electric aerial vehicle.” And it shares its name with a widely-beloved futuristic cartoon family: the Jetsons. All sorts of technology that used to exist only in cartoon form has made its way into being since the show launched in 1962, from jet packs to 3D printed food to smartwatches. Swedish startup Jetson Aero’s tiny electric aircraft, the Jetson One, is sort of like George Jetson’s “car”—except there’s only space for one person, there’s not a closed cabin, and you can’t press a button to be dropped out the bottom. You can, however, press a button to activate a ballistic parachute, but Jetson Aero is really hoping none of its customers will ever have to use this feature. The parachute is a last-resort option, built into the Jetson One along with several other redundancies for passenger safety. The vehicle runs on battery power, with eight electric motors, and is like a helicopter in that it takes off and lands vertically (though the fact that it has eight propellers makes it a “multicopter”). The fastest it goes is 63 miles per hour (102 kilometers per hour), so about the same as highway driving in or near an urban area. At 9.3 feet long by 8 feet tall by 3.4 feet wide, the Jetson One is quite small, at least as far as aircraft go, and weighs 198 pounds (90 kilograms). It can carry a passenger that weighs up to 210 pounds, though the lighter you are, the longer you can fly for; a pilot weighing 187 pounds can fly for 20 minutes before the vehicle’s batteries need recharging. In the US the aircraft is classified as “ultralight,” meaning you don’t need a pilot’s license to fly it. Given its compact footprint, the vehicle could take off right from owners’ driveways; the company encourages potential customers to “make your garden or terrace your private airport.” It certainly would be nice to be able to fly without the hassle of first traveling to an airport, nor the expense of the associated storage and usage fees (though if you’re buying one of these little winged toys, those expenses probably won’t concern you too much; the Jetson One goes for $92,000, which actually isn’t outrageous given that it’s basically a personal mini plane). The downside of making your garden your personal airport, though, is that if you lose control of the vehicle or have a shaky takeoff or landing, you could go crashing right through your home’s roof or wall. In the spirit of its widely-known compatriot company, Ikea, Jetson Aero delivers its aircraft to customers as a partially-assembled kit, accompanied by “detailed build instructions.” It’s a long shot from shelving unit to personal eVTOL, though, and letting customers assemble the vehicle themselves seems an odd choice given the risks associated with getting even one piece wrong. Customers don’t seem worried, though. The company has already sold its entire 2022 production run (which, to be fair, was only 12 units) and is now taking orders for delivery in 2023. Image Credit: Jetson Aero

Friend or Foe? Single Neurons in the Brain Control Social Interaction, Study Finds Oct 26, 2021

Show notes

Neurons live in a society, and scientists just found the ones that may allow us to thrive in our own society. Like humans, individual neurons are strikingly unique. Also like humans, they’re constantly in touch with each other. They hook up into neural circuit “friend groups,” break apart when things change, and rewire into new cliques. This flexibility lets their collective society (the brain) and their owners (us) learn about and adapt to an ever-changing world. To rebuild neural circuits, neurons constantly monitor the status of their neighbors through sinewy branches that sprout from a rotund body. These aren’t just passive phone lines. Dotted along each branch are little dual-function units called synapses, which allow neurons to both chat with a neuron partner and record previous conversations. Each synapse retains a “log” of past communications in its physical and molecular structure. By “consulting” this log, a neuron can passively determine whether or not to form a network with that particular neuron partner—or even set of partners—or to avoid any interactions in the near future. Apparently, neurons can do the same for us. This week, by listening to the electrical chatter of single neurons in rhesus macaque monkeys, scientists at Harvard, led by Dr. Ziv Williams, honed in on a peculiar subset of neurons that helps us tell friends from foes. The neurons, sprinkled across the frontal parts of the brain, are strikingly powerful. As their monkey hosts hung out for game night, the little computers tracked how each player behaved—were they more cooperative, or selfish? Over time, these “social agent identity cells” stealthily map out the entire group dynamic. Using this info, the monkeys can then decide whether to team up with other monkeys or shun them. It’s not just correlation. By modeling the electrical activity of these neurons, scientists were able to accurately determine any monkey’s past decisions—and mind-blowingly, predict its future ones, basically “mind-reading” the animal’s next move. When the team tampered with the neurons’ activity with a short burst of electrical zaps, the monkeys lost their social judgment. Like the new kid in school, they could no longer decide who to befriend. “In the frontal cortex, these neurons appear to be tuned for possible action from peers, representing them as communication partners, competitors, and collaborators,” wrote Dr. Julia Sliwa at Sorbonne Université in Paris, who was not involved in the study. While the results are from monkeys, they’re some of the first to connect the activity of individual neurons to an extremely complicated yet necessary aspect of our lives. These data “are major steps in identifying the neural mechanisms for maneuvering in complex social structure,” she said. My Brain’s View of You We often think of neurons and circuits as hardware components that represent us: our perception, memory, decisions, feelings. Yet large parts of the brain are dedicated to representing other people in our external world. One famous example is the “Jennifer Aniston neuron.” Back in 2005, an experiment showed that a single neuron in a person’s brain could react to a particular face—for example, Aniston’s. A lightbulb moment for neuroscience and computer vision, the study raised a brazen idea: that a single neuron has the computing power to encode a person’s physical identity. In the real world, it gets more complicated than just identifying a face. A person’s face comes with history—is it my first time meeting them? What’s their reputation? How do I feel about this person? Much work on how our brains handle social interaction comes from studying bands in music studios or teacher-student dynamics in classrooms. Here, brain activity is captured with wearables, which measure brain waves that wash over parts of the brain. These studies show that when making music or watching a movie together, our brain waves sync up. To rephrase, our brains tune into other peoples’—but when,...

Not So Mysterious After All: Researchers Show How to Crack AI’s Black Box Oct 25, 2021

Show notes

The deep learning neural networks at the heart of modern artificial intelligence are often described as “black boxes” whose inner workings are inscrutable. But new research calls that idea into question, with significant implications for privacy. Unlike traditional software whose functions are predetermined by a developer, neural networks learn how to process or analyze data by training on examples. They do this by continually adjusting the strength of the links between their many neurons. By the end of this process, the way they make decisions is tied up in a tangled network of connections that can be impossible to follow. As a result, it’s often assumed that even if you have access to the model itself, it’s more or less impossible to work out the data that the system was trained on. But a pair of recent papers have brought this assumption into question, according to MIT Technology Review, by showing that two very different techniques can be used to identify the data a model was trained on. This could have serious implications for AI systems trained on sensitive information like health records or financial data. The first approach takes aim at generative adversarial networks (GANs), the AI systems behind deepfake images. These systems are increasingly being used to create synthetic faces that are supposedly completely unrelated to real people. But researchers from the University of Caen Normandy in France showed that they could easily link generated faces from a popular model to real people whose data had been used to train the GAN. They did this by getting a second facial recognition model to compare the generated faces against training samples to spot if they shared the same identity. The images aren’t an exact match, as the GAN has modified them, but the researchers found multiple examples where generated faces were clearly linked to images in the training data. In a paper describing the research, they point out that in many cases the generated face is simply the original face in a different pose. While the approach is specific to face-generation GANs, the researchers point out that similar ideas could be applied to things like biometric data or medical images. Another, more general approach to reverse engineering neural nets could do that straight off the bat, though. A group from Nvidia has shown that they can infer the data the model was trained on without even seeing any examples of the trained data. They used an approach called model inversion, which effectively runs the neural net in reverse. This technique is often used to analyze neural networks, but using it to recover the input data had only been achieved on simple networks under very specific sets of assumptions. In a recent paper, the researchers described how they were able to scale the approach to large networks by splitting the problem up and carrying out inversions on each of the networks’ layers separately. With this approach, they were able to recreate training data images using nothing but the models themselves. While carrying out either attack is a complex process that requires intimate access to the model in question, both highlight the fact that AIs may not be the black boxes we thought they were, and determined attackers could extract potentially sensitive information from them. Given that it’s becoming increasingly easy to reverse engineer someone else’s model using your own AI, the requirement to have access to the neural network isn’t even that big of a barrier. The problem isn’t restricted to image-based algorithms. Last year, researchers from a consortium of tech companies and universities showed that they could extract news headlines, JavaScript code, and personally identifiable information from the large language model GPT-2. These issues are only going to become more pressing as AI systems push their way into sensitive areas like health, finance, and defense. There are some solutions on the horizon, such as differential privacy, where mode...

Our TOPPODCAST Picks

Follow Us

Stay Connected

Related Podcasts

Links

Stay Connected