AI with AI: Artificial Intelligence with Andy Ilachinski

Andy and Dave discuss the latest in AI news, including a report that systematically examined 62 studies on COVID-19 ML methods (from a pool o 2200+ studies), and found that none of the models were of potential clinical use due to methodological flaws or underlying biases. MIT and Amazon identify pervasive label errors in popular ML datasets (such as MNIST, CIFAR, Imagenet) and demonstrate that models may learn systematic patterns of label error in order to improve their accuracy. DARPA's Air Combat Evolution program upgrades its virtual program to include new weapons systems and multiple aircraft, with live Phase 2 tests on schedule for later in 2021. Researchers at the University of Waterloo and Northeastern University publish research working toward self-walking robotic exoskeletons. British researchers add a buccinators (cheek) muscle to robotic busts to better synchronize speech and mouth movements. Russian Promobot is developing hyper-realistic skin for humanoid robots. And Anderson Cooper takes a tour of Boston Dynamics. In research, Leverhulme, Cambridge, Imperial College London, and DeepMind UK publish research on the direct human-AI comparison in the animal-AI environment, using human children ages 6-10 and animal-AI agents across 10 levels of task groupings. Josh Bongard and Michael Levin publish Living Things Are Not (20th Century) Machines, a thought piece on updating how we think of machines and what they *could* be. Professors Jason Jones and Steven Skiena are publishing a running AI Dashboard on Public Opinion of AI. The Australian Department of Defence publishes A Method for Ethical AI in Defence. Raghavendra Gadagkar publishes Experiments in Animal Behavior. And Peter Singer and August Cole publish An Eye for a Storm, envisioning a future of professional military education for the Australian Defence Force.

Listeners Survey: https://bit.ly/3bqyiHk

Click here to visit our website and explore the links mentioned in the episode.

Andy and Dave discuss the latest in AI news, including a letter from the National Transportation Safety Board that asks the National Highway Traffic Safety Administration to regulate more strictly autonomous vehicles and driver assistance technologies; of note, the letter also uses Tesla as an example, stating that the company is using its customers to beta test its full self-driving technology on public roads. KMPG surveys business leaders on a variety of AI-related topics and finds that, among other things, many more leaders have the perception that AI tech is moving out too quickly. Researchers at Aston University announce a three-year study to explore the utility of human brain stem cells grown on a microchip, a so-called Neu-ChiP. Researchers from Norway and Australia unveil DyRET, a quadruped robot that can adapt its morphology (such as growing taller or shorter) as it encounters different environments. And Japanese researchers describe a decoded neurofeedback (DecNef) method, which uses fMRI to visualize brain activity and then calculate the similarity between real-time brain activity and brain activity patterns corresponding to specific pre-established memory and mental states. Microsoft's PowerPoint has a Presenter Coach that will listen and watch your presentation and give you pointers on speech patterns, pacing, attention, body language, and other attributes. The two main research items both involve AI agents playing in the Atari Learning Environment (57 games from Atari's library), and both with groundbreaking results in different ways: Uber AI and OpenAI use a model-free approach in Go-Explore, which uses a concept of "first return (to previous states), and then explore; GoogleAI use a world model approach with DreamerV2, which learns behaviors inside a separately trained world model (they also recommend a "clipped record mean" to aggregate scores across the various games). The survey of the week looks at Deepfakes Generation & Detection. Marjorie McShane and Sergei Nirenburg publish Linguistics for the Age of AI, arguing that researchers must place linguistics front and center for machines to achieve human-level language understanding, with big data and stats approaches as contributing methods. And in the video of the week, Steven Gouveia has produced a documentary on The Age of AI.

Listeners Survey: https://bit.ly/3bqyiHk

Click here to visit our website and explore the links mentioned in the episode.

Andy and Dave discuss the latest in AI news, including the release of the U.S. Navy and Marine Corps Unmanned Campaign Framework, which describes the desired approach to developing and deploying unmanned systems. Google employees demand stronger laws to protect AI researchers, in the wake of the firings of Gebru and Mitchell. Hour One debuted technology that creates fully digital and photorealistic AI personas for the purposes of content creation, such as welcome receptionist or information desk. Pennsylvania state law now allows for autonomous delivery robots to use sidewalks and operate on roads. The U.S. Army announces the availability of a training set for facial recognition that also includes thermal camera images, which it will make available for "valid scientific research." In research, Facebook AI demonstrates an algorithm capable of human-level performance in Diplomacy (no-press), using an equilibrium search to reason about what the other players are reasoning; the algorithm achieved a rank of 23 out of 1,128 human players. Researchers in Helsinki and Germany explore the effects of the Uncanny Valley, suggesting that a robot's appearance changes how humans judge its decisions. The Resource of the Week comes via Pete Skomoroch, who pointed out that Wikipedia contains a massive list of datasets for machine learning research (along with useful summary details about the dataset). The Book of the Week is Telling Stories, with authors from around the globe bringing culturally different perspectives on tales of AI. And the Videos of the Week come from MIT, which has published its Introduction to Deep Learning course online, with free access.

Listeners Survey: https://bit.ly/3bqyiHk

Click here to visit our website and explore the links mentioned in the episode.

Andy and Dave discuss the latest in AI news, including an announcement from Facebook AI that it achieved state of the art computer vision performance with its SEER model, by learning from one billion (with a 'b') random, unlabeled, and uncurated public Instagram images, reaching 84% top-1 accuracy on 13k images from ImageNet. DARPA launches a new Perceptually-enabled Task Guidance (PTG) to help humans perform complex tasks (such as through augmented reality); the effort will include both fundamental research as well as integrated demonstrations. DARPA also announces research teams from its Semantic Forensics (SemaFor) effort at probing media manipulations. Chris Ume, a Belgian visual effects artist, releases four deepfake videos of Tom Cruise, using two NVIDIA GPUs, two months training time, and further days of processing and tweaking for each clip. Researchers at the University of Washington, Berkeley, and Google Research use the StyleGAN2 framework to create "time-travel photography," which peels away the limitations of early cameras to reveal restored images of the original photos; the effort also involves the creation of a modern "sibling," which then gets merged with the original. OpenAI publishes the discovery that neurons in its CLIP network respond to the same concept, whether literal, symbolic (e.g., a sketch) , or conceptual (e.g., text); they also discover an absurdly simple attack, which involves places a stick with a word onto an item. The report of the week from UNICEF looks at Adolescent Perspectives on AI, with insights from 245 adolescents from five countries. Montreal.AI provides a 33-page "cheat sheet" with condensed information and links on AI topics. The book of the week from E-IR examines Remote Warfare: Interdisciplinary Perspectives. And the fun site of the week, MyHeritage, lets users animate photos, or "re-animate your dead loves ones."

Listeners Survey: https://bit.ly/3bqyiHk

Click here to visit our website and explore the links mentioned in the episode.

Andy and Dave discuss the latest AI news, which includes lots of new reports, starting with the release of the final report of the National Security Commission on AI, with over 750 pages that outlines steps the U.S. must take to use AI responsibly for national security and defense. The Stanford University Institute for Human-Centered AI (HAI) releases its fourth and most comprehensive report of its AI index, which covers global R&D, technical performance, education, and other topics in AI. Peter Layton at the Defence Research Centre in Australia publishes Fighting AI Battles: Operational Concepts for Future AI-Enabled Wars, with a look at war at sea, land, and air. Drone Wars in the UK and the Centre for War Studies in Denmark release Meaning-Less Human Control: Lessons from Air Defence Systems on Meaningful Human Control for the Debate of AWS, examining automation and autonomy in 28 air defense systems used around the world. And the European Union Agency for Cybersecurity publishes a report on Cybersecurity Challenges in the Uptake of AI in Autonomous Driving. In research, scientists demonstrate that an organism without a nervous systems, slime mold, can encode memory of its environment through the hierarchy of its own tube diameter structure. And the Fun Site of the Week uses GPT-3 to generate classic "title/description/question" thought experiments.

Listeners Survey: https://bit.ly/3bqyiHk

Click here to visit our website and explore the links mentioned in the episode.

Andy and Dave discuss the latest AI news including, Mars landing of the Perseverance and its AI-related capabilities, along with its mini-helicopter, Ingenuity. Researchers from Liverpool use machine learning to predict which mammalian hosts can generate novel coronaviruses. Researchers from Estonia and France create artificial human genomes using generative neural networks. A coalition of over 40 organizations have written a letter to ask that President Biden ban the federal use of and funding of facial recognition technology. The law firm Gibson Dunn releases a 2020 Annual Review of AI and Automated Systems, which also contains a great summary of policy and regulatory developments in the last year. In research, scientists at the Commonwealth Scientific and Industrial Research Organisation in Australia use AI to manipulate human behavior, steering participants toward particular actions. Researchers in the Netherlands demonstrate that predictive coding in recurrent neural networks naturally arises as a consequence of minimizing energy consumption. Research in Nature Communications demonstrates a multisensory neural networks that integrates information from all five human senses. The report of the week comes from CSET author Matthew Mittlelsteadt, which describes AI Verification: Mechanisms to Ensure AI Arms Control Compliance. The first book of the week comes from Moritz Hardt, on Patterns, Predictions, and Actions: A story about machine learning. And the fun site of the week takes a look at the works of painter Wassily Kandinsky, who was also a synesthete (experiencing the fusion of the senses), and offers insights into what he might have heard from looking at his paintings. The second book of the week provides some great information on Synaesthesia – Opinions and Perspectives.

Listeners Survey: https://bit.ly/3bqyiHk

Click here to visit our website and explore the links mentioned in the episode.

In AI news, researchers from the University of Copenhagen develop a machine learning model that estimates the chances of risk of death due to COVID at various stages of a hospital stay, including a 80 percent accuracy whether a patient with COVID will require a respirator. The Joint AI Center has a double-announcement, with the Tradewind Initiative, which seeks to develop an acquisition ecosystem to speed the delivery of AI capabilities, and with Blanket Purchase Agreements for AI testing and evaluation services. Kaggle publishes a survey on the 2020 State of Data Science and ML, which examines information from ~2000 data scientists about their jobs and their experiences. PeopleTec releases an "Overhead MNIST," a dataset containing benchmark satellite imagery for 10 categories (parking lots, cars, plans, storage tanks, and others). Epic's Unreal Engine introduces the MetaHuman Creator for release later this year, which purports to create ultra-realistic visuals for virtual human characters; Andy uses the moment to describe the "Uncanny Valley," which the Epic tech might manage to leap out of. And researchers from Carnegie Mellon and George Washington show that, like language transformers, image representations contain human-like biases. In research, researchers at the Israel Institute of Technology create a Ramanujan Machine, which can generate conjectures for mathematical constants, without proof. Researchers demonstrate initial steps of reconstructing video from brain activity. The report of the week examines U.S. public opinion on AI, with views on declining support for development and divided views on facial recognition. DeepMind London approaches the topic of common sense from the viewpoint of animals. And the book of the week comes from the author of the aforementioned paper, Murray Shanahan, and his 2010 book Embodiment and the Inner Life.

Listeners Survey: https://bit.ly/3bqyiHk

In news, Andy and Dave discuss a machine learning algorithm from Synergies Intelligent System and Universität Hamburg that can identify people in a moving crowd who are mostly likely asymptomatic carriers of COVID-19. US lawmakers have introduced the Public Health Emergency Privacy Act, to boost privacy protections for COVID-19 technology such as tracing apps and vaccine scheduling apps. A team led by researchers from Oxford have introduced new reporting guidelines to bridge a gap in development to implementation when using clinical AI technologies, dubbed DECIDE-AI. Over 30 authors from a wide swath of organizations have proposed a "living benchmark" to evaluate progress in natural language generation, which they call GEM (Generation, Evaluation, and Metrics). And the combination we saw coming, research from Queen Mary University demonstrate a deep learning framework for detection of emotion using wireless signals. Researchers at the University of Virginia claim to detect physiological responses to racial bias with 76.1% accuracy, though it more focuses on exploring any link between mental associations of skin color. In research, Stanford researchers explore how learning and evolution occur in complex environments, and how they affect the diversity of morphological forms, with DERL (Deep Evolutionary Reinforcement Learning). Researchers from University of Illinois, Urbana-Champaign, introduce GANs for editing images via their latent space, which provides greater control over editing (e.g., editing a mouth without re-generating the entire face). And in the video of the week, a 12-minute video provides a short history on DARPA with highlights on many of its military robot programs. Listener Survey: https://bit.ly/3bqyiHk

In COVID-related AI news, Andy and Dave discuss research from Texas &AM, Wisconsin-Milwaukee, and SNY Binghamton, which demonstrates an automatic system for monitoring the physical distance and face mask wearing of construction workers; demonstrating how surveillance is rapidly becoming a widely available commodity technology. In regular news, the National Security Commission on AI releases its draft final report, which makes sweeping recommendations on AI as a constellation of technologies. The nominee for Deputy Secretary of Defense, Kathleen Hicks, mentions AI and the JAIC at several points during her testimony. The Information Technology & Innovation Foundation releases a report on "Who Is Winning the AI Race," using 30 different metrics to assess nations' progress in AI. Amnesty International launches a campaign against facial recognition, dubbed "Ban the Scan." And Scatter Lab pulls its Korean chatbot Lee Luda, after it started responding with racist and sexist comments to user inputs. In three "quick" research items, researchers at Massachusetts General Hospital and Harvard Medical School show that single neurons can encode information about others' beliefs. Researchers at MIT and the Institute of Science and Technology Austria introduce a new class of time-continuous recurrent neural network models, which they dub liquid time-constant networks; the approach reduces the size of networks by nearly two orders of magnitude for some tasks. And researchers at the University of Toronto, Microsoft Research, and Cornell University show that Maia, a custom version of AlphaZero, can learn to predict human actions, rather than the most likely winning move. The report of the week looks at The Immigration Preferences of Top AI Researchers. And the book of the week contains almost 40 chapters and 60 authors on a variety of special operations-related topics, in Strategic Latency Unleashed.

Listener Survey

Listener Survey In COVID-related AI news, Andy and Dave discuss a machine learning transformer model from Facebook AI and the NYU School of Medicine that uses x-rays to determine whether a COVID patient might need more intensive care. A for-pay report from Synced provides a survey of China's AI efforts in response to COVID-19. In regular AI news, European Parliament members adopt guidelines for military and non-military uses of AI. Meanwhile, the UK Competition and Markets Authority cautions that algorithms can damage online competition and should face regulatory scrutiny. Researchers at NOIRLab use machine learning to identify just over 1200 potential new gravitational lenses. Researchers at Harvard use fish-inspired robots to demonstrate coordinated swarm movements without any outside control. Nature provides reflections from various authors on AI. And the AI Newsletter compiles a list of the 100 most influential people in AI. In research topics, researchers at Cornell demonstrate a curriculum strategy to solve hard Sokoban (the "warehouse man" game) problems, and builds on a pool of sub-tasks. And in a similar, but unrelated effort, researchers at Berkeley and Google Research create a trio of agents to create challenging but feasible environments for the primary agent, the protagonist to navigate; they use an antagonist agent, which tries to create hard environments, while a third agent maximizes the differential between the other two agents, which keeps the tasks just at the edge of the protagonist's ability to solve. An article in the Journal of AI Research demonstrates that containment of a superintelligence is impossible, due to fundamental limits inherent in computing itself. And finally, Chitta Ranjan provides the book of the week, in Understanding Deep Learning: Application to Rare Event Prediction.

Click here to visit our website and explore the links mentioned in the episode.