AI with AI: Artificial Intelligence with Andy Ilachinski

Happy Pi-cast! Andy and Dave discuss some of the stories that have followed the New York Times articles on Clearview AI, to include Twitter telling the company to stop using its photos, and a consortium of 40 agencies calls on the U.S. government to ban facial recognition systems until more is known about the technology. Meanwhile, London’s Metropolitan Police is rolling out live facial recognition technology. BlueDot says that it used AI and its epidemiologists to send a warning about the Wuhan virus on 31 December 2019, a full week before the US CDC announcement on 6 January 2020. Google releases the largest high-resolution map of the fruit fly’s brain, with 25,000 neurons. DARPA’s Gremlin (X-61A) drone system makes its first test flight. And the Guinness Book of World Records recognizes Stephen Worswick as the most frequent winner (5 times) of the Loebner Prize, for his Mitsuku chatbot. In research, Facebook AI achieves near-perfect (99.9%) navigation without needing a map, testing its algorithm in its AI Habitat. Robert J. Marks makes The Case *for* Killer Robots. The Brookings Institute’s Indermit Gill predicts that the AI leader in 2030 will “rule the planet” until at least 2100. The ACT-IAC releases an AI Playbook, with step-by-step guidance for assessment, readiness, selection, implementation, and integration. Jessica Flack examines the Collective Computation of Reality in Nature and Society. Google’s Dataset Search is out of beta. And DoD will be holding its East Coast AI Symposium and Exposition 29 and 30 April in Crystal City.

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

In a string of related news items on facial recognition, Andy and Dave discuss San Diego’s reported experiences with facial recognition over the last 7 years (coming to an end on 1 January 2020 with the enacting of California’s ban on facial recognition for law enforcement).  Across the Atlantic, the European Union is considering a ban on facial recognition in public spaces for 5 years while it determines the broader implications. And the New York Times puts the spotlight on Clearview AI, a company that claims to have billions of photos of people scraped from the web, and that identify people (and the sources of the photos, to include profiles and other information about the individuals) within seconds. In other news, the JAIC is looking for public input on an upcoming AI study, and it is also looking for help in applying machine learning to humanitarian assistance and disaster relief efforts. In research, Google announces that it has developed a “physics free” model for short-term local precipitation forecasting. And researchers at DeepMind and Harvard find experimental evidence that dopamine neurons in the brain may predict rewards in a distributional way (with insight gained from efforts in optimizing reinforcement-learning algorithms). Nature Communications examines the role of AI, whether positive or negative, in achieving the United Nations’ Sustainable Development Goals. The U.S. National Science Board releases its biennial report on Science and Engineering Indicators. The MIT Deep Learning Series has Lex Fridman speaking on Deep Learning State of the Art (and as a bonus, Andy recommends a video of Fridman interviewing Daniel Kahneman, author of “Thinking, Fast and Slow”). GPT-2 wields its sword and dashed bravely into the realm of Dungeons and Dragons. And GPT-2 tries its hand at chess, knowing nothing about the rules, with surprising results.

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

The U.S. Government announces the restriction of the sale outside of the U.S. of AI for satellite image analysis. Baidu beats out Google and Microsoft for language “understanding” with its model ERNIE, which uses a technique that it developed specifically for the Chinese language. Samsung unveils NEON, its humanoid AI avatars. The U.S. Department of Defense stands up a counter-unmanned aerial system office. And GoogleAI publishes an AI system for breast cancer screening, but meets with some Twitter (and Wired) backlash on solving the “wrong problem.” Researchers at University of Vermont, the Allen Discovery Center/Tufts, and Wyss Institute/Harvard introduce the world’s “first living robots,” xenobots, constructed from skin and muscle cells of frogs (from designs made with evolutionary algorithms). RAND releases a report on an assessment and recommendations of the DOD’s posture for AI. AI for social good (AI4SG) releases its survey of research and publications on beneficial applications of AI. Daniel Dennett explores the question of whether HAL committed murder, in a classic 1996 essay. From the Bengio and Marcus debate, both references Daniel Kahneman’s “Thinking, Fast and Slow.” And Robert Downey Jr. hosts a YouTube series on The Age of AI.

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

Andy and Dave discuss a new White House proposal on Principles for AI Regulation. A NIST study examines the effects of race, age, and sex on recognition software and identifies a variety of troubling issues. Facebook removes hundreds of accounts with AI-generated fake profile photos, and Facebook also bans the posting of deepfake videos (with some caveats). And Finland is making its online AI course available for the rest of the world. In research, Uber AI Labs offers a novel approach to accelerating neural architectural search by learning to generate synthetic training data; but the scientific community doesn’t think the findings are quite yet ready for publishing. Researchers at Korea University create an Evolvable Neural Unit (ENU) as a way to approximate the function of an individual neuron and synapse. And researchers at Charite in Berlin show that a single human biological neuron can compute XOR, previously thought not possible. Human-Centered AI at Standard University releases 2019 Annual Report on its AI Index, examining various trends and research in AI in 2019. The Center for a New American Security releases its full report on A Blueprint for Action in AI. Rafael Irizarry provides an Introduction to Data Science. And the video of the week is the debate between Yoshua Bengio and Gary Marcus on the current and future state of research in AI.

In research, Andy and Dave discuss a new idea from Schmidhuber, which introduces Upside-Down reinforcement learning, where no value functions or policy search are necessary, essentially transforming reinforcement learning into a form of supervised learning. Research from OpenAI demonstrates a “double-descent” inherent in deep learning tasks, where performance initially gets worse and then gets better as the model increases in size. Tortoise Media provides yet-another-AI-index, but with a nifty GUI for exploration. August Cole explores a future conflict with Arctic Night. And Richard Feynman provides thoughts (from 1985) on whether machines will be able to think.

Twitter Throwdown: On 23 December, Yoshua Bengio and Gary Marcus will have a debate on the Best Way Forward for AI.

Andy and Dave discuss Lee Sodol’s announcement that he is quitting playing Go because AI “cannot be defeated.” Facebook’s Head of AI says the field will soon “hit the wall” (or does he?). A human beats an AI-powered drone during the Drone Racing League’s latest competition. A Boston Dynamics robot dog has joined a Massachusetts bomb squad. And a new US federal bill would constrain some police use of facial recognition tools. A report from CNAS on the American AI Century provides a Blueprint for Action on how to achieve national AI strategy objectives. 

The White House Office of Science and Technology Policy releases its 2016-2019 Progress Report on Advancing AI R&D. SkyNet publishes “AI Coverage Best Practices, According to AI Researchers.” The US Patent and Trademark Office asks whether AI can own the content that it creates. And Amazon files paperwork to protest the awarding of Pentagon’s JEDI cloud project o Microsoft. DeepMind announces the latest iteration of its board game playing algorithm, with MuZero, which learns to play go, chess, and shogi (and Atari games) without any knowledge of the game rules. Researchers from Duke publish a deep learning model for interpretable image recognition. Bill Hibbard pens thoughts on Ethical AI. A review paper examines the evolution of explainable AI (XAI) and interpretable AI. Mick Ryan looks at Extending the Intellectual Edge with AI. Melanie Mitchell has a video redux of her recent book AI: A Guide for Thinking Humans. And Google releases Version 2.0 of its Teachable Machine.

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

Andy and Dave discuss OpenAI’s update to an earlier summary of how computational resources have increased to reach each new AI breakthrough. The National Transportation Safety Board releases its report on the 2018 deadly Uber self-driving vehicle crash. Nasdaq has enlisted the aid of machine learning to provide additional security to stock trades. Researchers use a GAN to GANalyze the aspects of “memorable” pictures, while other researchers use a GAN (SinGAN) to generate new pictures from a single image. Over 20 authors come together to publish a paper on tackling climate change with machine learning. Francois Chollet publishes The Measure of Intelligence. Horace He has provided OpenReviewExplorer to include the International Conference on Learning Representations 2020. And FRONTLINE examines the promise and perils of AI.

Andy and Dave discuss the full release of the algorithm that originally had to be locked up for the safety of humanity (GPT-2). NATO releases its final reports on the implications of AI for NATO’s Armed Forces. The US Army Research Lab wraps up a series of events on its efforts in robotics collaborative technology. The UAE announces the world’s first graduates level AI University opening in September 2020. And John Carmack announces he will step down at CTO of Oculus to tackle the challenge of artificial general intelligence, as a Victorian Gentleman Scientist. In research, two independent research groups introduce adversarial T-shirts. A report examines the taxonomy of real faults in deep learning systems. Krohn, Deyleveld, and Bassens publish Deep Learning Illustrated. The Nov/Dec issue of MIT Technology Review features a variety of AI and related stories. And Manuel Blum of CMU discusses Towards a Conscious AI: A Computer Architecture Inspired by Neuroscience.

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

In news, the Defense Innovation Board releases AI Principles: Recommendations on the Ethical Use of AI by the Department of Defense. The National Institute of Standards and Technology’s National Cybersecurity Center of Excellent releases a draft for public comment on adversarial machine learning, which includes an in-depth taxonomy on the possibilities. Google adds BERT to its search algorithm, with its capability for bidirectional representations, in an attempt to “let go of some of your keyword-ese.” In research, Stanford University and Google demonstrate a method for explaining how image classifiers make their decisions, with Automatic Concept-based Explanations (ACE) that extra visual concepts such as colors and textures, or objects and parts. And GoogleAI, Stanford, and Columbia researchers teach a robot arm the concept of assembling objects, with Form2Fit, which is also capable of generalizing its learning to new objects and tasks. Danielle Tarraf pens the latest response to the National Security Commission on AI’s call for ideas, with Our Future Lies in Making AI Robust and Verifiable. Jure Leskovec, Anand Rajaraman, and Jeff Ullman make their second edition of Mining of Massive Datasets available. The Defense Innovation Board posts a video of its public meeting from 31 October at Georgetown University. Maciej Ceglowski’s “Superintelligence: the idea that eats smart people” takes a look at the arguments against superintelligence as a risk to humanity.

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