AI with AI: Artificial Intelligence with Andy Ilachinski

Andy and Dave discuss the Joint Artificial Intelligence Center’s efforts to tackle deep fakes through DARPA’s Media Forensics program, as well as the announcement that the JAIC’s biggest project for FY20 will include “AI for maneuver and fires.” Intel reveals its first AI chips, on the Nervana Neural Network Processor line, with one to train AI systems and another to handle inference. Cerebras Systems announces the world’s largest chip, with 1.2 trillion transistors and 400,000 cores. A Russian Soyuz spacecraft docked with the International Space Station; it had Roscosmos’s Skybot F-850 humanoid robot aboard. Researchers at Hong Kong University of S&T demonstrate an all-optical neural network for deep learning. Researchers at MIT and Tubingen identify four types of neuronal cells based on their electrical spiking activity. And a larger team of researchers, primarily based in China, unveil the Tianjic chip, as a hybrid that combines computer science (with a binary focus) with neuroscience (with a neural burst and spike focus) on one chip. In the book of the week, K. Eric Drexler of Oxford publishes a large report on Reframing Superintelligence. An article from Melanie Mitchel in Popular Computing in 1985 seems hardly out of place in 2019 with its look at what people were predicting for the future. A report from PAX surveys the tech sector’s stance on lethal autonomous weapons. The Intelligence Community Studies Board releases the proceedings of a workshop on Robust Machine Learning Algorithms and Systems for Detection and Mitigation of Adversarial Attacks and Anomalies. Jonathan Clifford pens a piece in War on the Rocks on how “AI will change war, but not in the way you think.” In a video, Elon Musk and Jack Ma discuss AI at the World AI Conference in Shanghai. And the Australian Defence College will host a seminar on Science Fiction and the Future of War on 3 October 2019.

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

Happy 100th Episode to AI with AI! Andy and Dave celebrate the 100th episode of the AI with AI podcast, starting with a new theme song, inspired by the Mega Man series of games. Andy and Dave take the time to look at the past two years of covering AI news and research, including at how the podcast has grown from the first season to the second season. They also take a look back at some of the recurring themes and favorite topics, including GPT2 and the Lottery Ticket hypothesis, among many others; they also look forward to (hopefully!) all the latest and greatest news to come. Throughout this episode, we hear from listeners, supporters, and colleagues who have appeared on the podcast. Here’s to another 100, and thanks for listening!

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

Researchers at Berkeley, Washington, and Chicago identify “natural adversarial” examples that cause classifier accuracy to significantly degrade, likely due to an over-reliance on color, texture, and background cues. Andy and Dave then discuss a series of events following a Nature paper on application of deep learning to aftershock patterns of earthquakes, wherein other researchers raised questions on the researcher (one demonstrating that a simple logistic regression does better; and another showing that the original researchers included their test data set in their training data set). A new study by the Insurance Institute for Highway Safety shows that drivers overestimate the capability of vehicle automated systems, with Telsa’s Autopilot leading the rest in overestimation. Goodfellow, Bengio, and Courville publish their 800 page tome on Deep Learning. The Classic Paper of the Week comes from Pattie Maes and Rodney Brooks, who published Learning to Coordinate Behaviors in 1990. The video presentation of the octopus research makes the video of the week. NASA streams 24/7 with OUTERHELIOS, a neural network trained on Coltrane to produce non-stop free jazz (though the feed may now be “static only”).

Click here to leave a testimonial for the 100th episode!

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

The National Security Commission on AI solicits creative and original ideas to challenge the status quo assumptions on maintaining US global leadership in AI. Researchers at MIT and Colgate publish an engineering *concept* that would use superconducting nanowires to mimic artificial neurons in a way that would theoretically match the energy efficiency of brains. Microsoft invests $1B in OpenAI to create brain-like machines. A proposed bill would prohibit the use of facial recognition technology for all public housing units that receive funding from the Department of Housing and Urban Development. Researchers at the University of Washington, Seattle demonstrate that octopuses' arms are capable of making decisions without input from their brains, with more than 350 million of its 500 million neurons in their arms. Google DeepMInd uses a generative adversarial model to generate fictional videos with DVD-GAN.

Click here to leave a testimonial for the 100th episode!

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

The University of Singapore creatures an artificial skin that can sense temperature, pressure, and humidity. The International Center for Ethics in the Sciences and Humanities releases its Evaluation of (AI) Guidelines. A report from FutureGrasp takes a global look at the AI initiatives (or lack thereof) of States. Hayden Klok and Yoni Nazarathy release a draft of Statistics with Julia. Metaacademy provides learning plans and resources for learning about topics, from beginner to advanced. Claude Shannon’s 1948 paper “A Mathematical Theory of Communication” makes Andy’s Classic Paper for the week. Stephen Wolfram’s testimony on AI before the US Senate Commerce Committee becomes available, including his blog write-up about the testimony. And Fedor Kitashov publishes an essay on using AI to restore and colorize photos.

Click here to leave a testimonial for the 100th episode!

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

Andy and Dave discuss the Digital Modernization Strategy that the US Department of Defense released on 12 July 2019. Todd Austin at the University of Michigan presents research at a conference on Morpheus, a project to create a chip that randomizes elements of its code, in an attempt to slow would-be hackers. Also in chip-related news, Intel introduces Pohoiki Beach, a new 8 million-neuron neuromorphic system with 64 Loihi research chips, with expectations that they will produce a system capable of simulation 100 million neurons by the end of 2019. Baylor College of Medicine in collaboration with the University of California and Second Sight Medical Products announce Project Orion, an implant that transmits video images directly to the visual cortex, bypassing the eye and optic nerve.  And the Naval Information Warfare Systems Command and PEO C4I announce the AI Applications to Autonomous Cybersecurity (AI ATAC), a contest for using AI/ML to bolster network security operations. Research from University of Wisconsin, Madison, demonstrates that optical waves passing through a nanophotonic medium can perform artificial neural computing – here, that a sheet of glass can identify numbers by “looking,” or in this case, by making use of bubbles and other impurities in the glass to function as a neural processor.  Research from Stanford creates a convolutional neural network that can play Go without game tree search, more mimicking a human-level understanding and approach.

Click here to leave a testimonial for the 100th episode!

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

Continuing in research, Andy and Dave discuss research from Imperial College and the Samsung AI Centre, which can take a single image of any face, and create realistic speech-driven facial animations, using a GAN. From the Conference on Computer Vision and Pattern Recognition, researchers create an algorithm that can learn individual styles of conversational gesture, and then produce plausible gestures to accompany other audio input. And research in Nature examines 3.3 million material-science abstracts with unsupervised word embeddings to capture “latent knowledge.” The survey paper of the week looks at the reproducibility of machine learning in health-related fields, and finds health consistently lags behind other subfield of machine learning. Safety First for Automated Driving identifies the guiding principles for autonomous cars to be safe, with input from 11 authors; among the information, the report finds that verification and validation of the systems is still lacking in the existing literature. The Berkman Klein Center at Harvard compiles an infographic on all of the published AI “principles” from governments, industry, and other organizations. The “classic paper” of the week comes from Alan Turing’s 1948 paper on “Intelligent Machinery.” The 36th International Conference on Machine Learning releases over 150 videos from its June session. CognitionX 2019 releases a video on managing security in an insecure world. Manlio de Domenico and Hiroki Sayama (and many others!) provide an interactive site for explaining and exploring complexity. Wendy Anderson and August Cole explore what war in the late 2020s might look like for the Secretary of Defense, in The Secretary of Hyperwar. And for click-bait of the week, astrophysicists get “baffled” by their simulation of the universe using AI.

Andy and Dave discuss a scathing report on Scotland Yard’s facial recognition software, which researchers at the University of Essex found to have an 81% error rate (but that the Met Police say has an error rate of 0.1%). In related news, Axon announced that it will ban the use of facial recognition systems on its devices; Axon supplies 47 of the 69 largest police agencies in the U.S. with body cameras and software. DARPA announces IDAS, the Intent-Defined Adaptive Software (IDAS), in an attempt to reduce the need for manual software modifications. NIST posts the first draft guideline for developing AI technical standards. Elon Musk says that its Neuralink is almost ready for the first human volunteers; Neuralink uses ultrafine threads that can be implanted into the brain to detect the activity of neurons. And the Bank of England announced that Alan Turing will be on the new Fifty Pounds note. In research, Andy and Dave discuss Pluribus, the latest AI for multiplayer poker from CMU and Facebook AI, which won during a 12-day poker marathon in 6-player no-limit Texas hold’em; the AI runs on two Intel processors and a “modest” 128GB during play.

More research from Berkeley and also University of Southern California creates a method to “protect” world leaders against deep fakes, by identifying, among other things, 17 Facial Action Units (such as subtle movements of eyebrows, cheeks, nose, etc, during speech). And research from MIT can take an audio clip and convert it to a generic human face. A report from RAND looks at Ethics in Scientific Research. Deakin University and Harvard provide a survey of deep reinforcement learning in cyber security. Another survey from Dublin University and Intel Labs looks at Generative Adversarial Networks and their taxonomy. Vishal Maini and Samer Sabri provide Machine Learning for Humans. Andy recommends Ludwig von Bertalanffy’s General System Theory from 1968. Matt Turek takes a look at the history of media forensics. The House Homeland Security Subcommittee on Intelligence and Counterterrorism holds a hearing on AI and Counterterrorism. And the Computer Vision and Pattern Recognition 2019 conference begins to post its tutorials, workshops, and its 80 page program guide.

Andy and Dave discuss the update to the US National AI Research and Development Strategic Plan, which establishes 8 objectives for federally funded AI research. Meanwhile, the European Commission starts its pilot phase for ethics guidelines for trustworthy AI, with the first AI Alliance Assembly meeting in Brussels and the High-Level Expert Group of AI (AI HLEG). The Joint AI Center, in conjunction with CMU, CrowdAI, and DIU, plans to make available xBD (x-Building-Damage), an open-source labeled data set of satellite imagery of some of the largest natural disasters in the past decade; it will contain ~700k building annotations across over 5,000 km^2 of imagery from 15 countries. The JAIC also announced a partnership with the Singapore’s Defence Science and Technology Agency to collaborate on AI in humanitarian assistance and disaster relief. A white paper by Pactera suggests that 85% of AI projects fail. A new DARPA program, Virtual Intelligence Processing (VIP) aims to explore “brain-inspired” methods for dealing with incomplete, spare, and noisy data. Facebook releases AI Habitat, an open source environment for training and testing AI agents. And NIST’s RFI on AI Standards receives nearly 100 respondents. Researchers at Adobe Research and Berkeley use AI to detect facial image manipulations that were done by Photoshop’s “Face Aware Liquify” feature; while humans were able to judge an altered face 53% of the time, the Convolutional Neural Network tool achieved results as high as 99%.

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