AI with AI: Artificial Intelligence with Andy Ilachinski

CNA's Center for Autonomy and Artificial Intelligence kicks off its first panel for 2019 with a live recording of AI with AI! Andy and Dave take a step back and look at the broader trends of research and announcements involving AI and machine learning, including: a summary of historical events and issues; the myths and hype, looking at expectations, buzzwords, and reality; hits and misses (and more hype!), and some of the many challenges of why AI is far from a panacea.

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

With Season 3 drawing to a close, Andy and Dave decided to focus this discussion entirely on the latest results from DARPA's Air Combat Evolution (ACE) program. On 20 August, DARPA held a contest between 8 competitors, and pitted their AI agents in simulated combat against each other, and against a human pilot (who used a VR system). Heron Systems won the event, beating out the other AI agents, and also not allowing the human pilot to attain a valid targeting solution. Andy and Dave discuss the results, the limitations, and the broader context of these results in light of other research and announcements.

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

In COVID-related AI news, Andy and Dave discuss survey from Amazon Web Surveys that examines the current status of Internet of Things applications related to COVID-19, include scenarios that might help to reduce the severity of an outbreak. MIT publishes an combinatorial machine learning method to maximize the coverage of a COVID-19 vaccine. In "quick takes" on research, Andy and Dave discuss research from Microsoft, University of Washington, and UC Irvine, which provides a checklist to help identify bugs in natural language processing algorithms. A paper from Element AI and Stanford examines whether benchmarks for natural language systems actually correspond to how we use those systems. University of Illinois at Urbana-Champaign, Columbia University, and US Army Research Lab introduce GAIA, which processes unstructured and heterogeneous multimedia data and creates a coherent knowledge base, and allows for text queries. Research published in Nature Neuroscience examines the brain connectivity of 130 mammalian species and finds efficiency of information transfer through the brain does not depend on the size or structure of any specific brain. And finally, Andy and Dave spend some time talking about the broader implications of GPT-3, the experiments that people are conducting with it, and how it is not an AGI. Dave concludes with an analogy from Star Trek: the Next Generation, that he gets mostly correct, though he misattributes Geordi La Forge's action to Dr. Pulaski. If only he had a positronic matrix!

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

In COVID-related AI news, Andy and Dave discuss an article from Wired that describes how COVID confounded most predictive models (such as finance). And NIST investigates the effect of face masks on facial recognition software. In regular-AI news, CSET and the Bipartisan Policy Center release a report on "AI and National Security," the first of four "meant to be a roadmap for Washington's future efforts on AI." The Intelligence Community releases its AI Ethics Principles and AI Ethics Framework. Researchers from the University of Chicago announce "Fawkes," a way to "cloak" images and befuddle facial recognition software. In research, OpenAI demonstrates that GPT-2, a generator designed for text, can also generate pixels (instead of words) to fill out 2D pictures. Researchers at Texas A&M, University of S&T of China, and MIT-IBM Watson AI Lab create a 3D adversarial logo to cloak people from facial recognition. And other research explores how the brain rewires when given an additional thumb. CSET publishes a Deepfakes: a Grounded Threat Assessment. And MyHeritage provides a "photo enhancer" that uses machine learning to restore old photos.

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

In COVID-related AI news, Tencent AI Labs publishes a "machine learning" model that can predict the risk of a coronavirus patient developing severe illness. Unsupervised machine learning on data from the U.K.'s COVID Symptom Tracker, which has more than 4 million users, suggests patients cluster into roughly 6 different symptom types. Amazon Web Services releases its version of a scientific literature search on COVID-19. Aminer.org offers an open access knowledge graph of COVID-19. And "Digital Contact Tracing for Pandemic Response" takes a look at global approaches and results with implementing contact tracing. In regular AI news, the National Security Commission on AI releases its latest quarterly report, with 35 recommendations. The latest Congressional Research Service Report covers Emerging Military Technologies, including AI and LAWS. Facebook rolls out a "bot army" to simulate "bad behavior" on a parallel version of its platform, in an effort to understand and combat online abuse. In research, DeepMind publishes findings on reinforcement learning, with a meta-learning approach that discovers an update rule that includes "what to predict" as well as "how to learn from it." Research from Berkeley, DeepMind, and MIT explores exploration by comparing how children learn with reinforcement learning agents in a unified environment. Military Review publishes an article by Courtney Crosby, which describes a framework for operationalizing AI for algorithmic warfare. DeepMind and University College London examines deep reinforcement learning and its implications for neuroscience. And MIT makes available online a full lecture series by Marvin Minsky on "The Society of Mind."

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

In COVID-related AI news, Andy and Dave discuss research that provides a comprehensive survey on applications of AI in fighting COVID-19. The Stanford Institute for Human-Centered AI and the AI Initiative at the Future Society launch a global alliance: Collective and Augmented Intelligence against COVID-19 (CAIAC). MIT and the IBM Watson AI Lab publish a paper that suggests a computational limit to progress in deep learning. The Atlas of Surveillance provides an open-source look at technologies that law enforcement are using across the US, to include facial recognition and drones. Similarly, Surfshark has compiled information on the status of facial recognition technology around the globe, along with additional useful information. MIT finds systematic shortcomings in the ImageNet dataset, with an observation that the crowdsourcing data collection pipeline can cause "misalignments." Research from Google Brain shows that "self-attention" can allow agents to identify task-critical visual hints, and ignore task-irrelevant elements. UC Berkeley, Google, CMU, and Facebook demonstrate "one policy to rule them all," where they use one global policy to control the movement of a wide variety of agent morphologies (which would normally require training and tuning for each separate morphology). The Army's Cyber Institute releases the "Invisible Force" graphic novel, which examines potential uses of AI technology in a future fictional scenario. Alife 2020 makes a compilation of its July conference available, clocking in a nearly 800 pages. And Gwern examines the creative side of GPT-3 through poetry, humor, and other probing interactions

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

Andy and Dave start with COVID-related AI news, and efforts from the Roche Data Science Coalition for UNCOVER (the United Network for COVID-19 Data Exploration and Research), which includes a dataset of a curated collection of over 200 publicly available COVID-19 related datasets; efforts from Akai Kaeru are included. The Biomedical Engineering Society publishes an overview of emerging technologies to combat COVID-19. Zetane Systems uses machine learning to search the DrugVirus database and information from the National Center for Biotechnology to identify existing drugs that might be effective against COVID. And researchers at the Walter Reed Army Institute of Research are using machine learning to narrow down a space of 41 million compounds to identify candidates for further testing. And the IEEE hosted a conference on 9 July, "Does your COVID-19 tracing app follow you forever?" In non-COVID-related AI news, MIT takes offline the TinyImages dataset, due to its inclusion of derogatory terms and images. The second (actually first) wrongful arrest from facial recognition technology (again by the Detroit Police Department) comes to light. Appen Limited releases its annual "State of AI and ML" report, with a look at how businesses are (or aren't) considering AI technologies. Anaconda releases its 2020 State of Data Science survey results. And the International Baccalaureate Educational Foundation turn to machine learning algorithms to predict student grades, due to COVID-related cancelations of actual testing, and much to the frustration of numerous students and parents. Research from the Vector Institute and the University of Toronto tackles analogy and the Raven Progressive Matrices with an ensemble of three neural networks for objects, attributes, and relationships. Researchers at the University of Sydney and the Imperial College London have established CompEngine, a collection of time-series data (over 24,000 initially) from a variety of fields, and have placed them into a common feature space; CompEngine then self-organizes the information based on empirical properties. Garfinkel, Shevtsov, and Guo make Modeling Life available for free. Meanwhile, Russell and Norvig release the not-so-free 4th Edition of AI: A Modern Approach. Lex Fridman interviews Norvig in a video podcast. And the Elias Henriksen creates the Computer Prophet, which generates metaphors from a database of collected sayings.

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

In COVID-related AI news, Purdue University has built a website that tracks global response to social distancing, by pulling live footage and images from over 30,000 cameras in 100 countries. Simon Fong, Nilanjan Dey, and Jyotismita Chaki have published Artificial Intelligence for Coronavirus Outbreak, which examines AI's contribution to combating COVID-19. Researchers at Harvard and Boston Children's Hospital use a "regular" Bayesian model to identify COVID-19 hotspots over 14 days before they occur. In non-COVID AI news, the acting director of the JAIC announces a shift to enabling joint warfighting operations. The DoD Inspector General releases an Audit of Governance and Protection of DoD AI Data and Technology, which reveals a variety of gaps and weaknesses in AI governance across DoD. Detroit Police Chief James Craig reveals that the police department's experience with facial recognition technology resulted in misidentified people about 96% of the time. Over 1400 mathematicians sign and deliver a letter to the American Mathematical Society, urging researchers to stop working on predictive-policing algorithms. DARPA awards the Meritorious Public Service Medal to Professor Hava Siegelmann for her creation and research in the Lifelong Learning Machines Program. And Horace Barlow, one of the founders of modern visual neuroscience, passed away on 5 July at the age of 98. In research, Udrescu and Tegmark release AI Feynman 2.0, with unsupervised learning of equations of motion by viewing objects in raw and unlabeled video. Researchers at CSAIL, NVidia, and the University of Toronto create the Visual Causal Discovery Network, which learns to recognize underlying dependency structures for simulated fabrics, such as shirts, pants, and towels. In reports, the Montreal AI Ethics Institute publishes its State of AI Ethics. In the video of the week, Max Tegmark discusses the previously mentioned research on equations of motion, and also discusses progress in symbolic regression. And GanBreeder upgrades to ArtBreeder, which can create realistic-looking images from paintings, cartoons, or just about anything.

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

It's a week of huge announcements! But first, in COVID-related AI news, Andy and Dave discuss a review paper in Chaos, Solitons, and Fractals that provides a more international focus on the role of AI and ML in COVID research. CSAIL teams with Ava Robotics to design a robot that maneuver between waypoints and disinfect surfaces of warehouses with UV-C light. C3.ai Digital Transformation Institute awards $5.4M to 26 AI researchers for projects related to COVID-19. In non-COVID news, the Association for Computing Machinery calls for the immediate suspension of facial recognition technologies until more mature and reliable. US lawmakers have introduced a bill that would ban police use of facial recognition, while separate bills seek to increase the AI talent available for the Department of Defense, and work to realign and rewire the JAIC within DoD. Over 2300 researchers sign a petition to Springer Nature to reject a publication from Harrisburg University, which developed facial recognition software to predict whether somebody was going to be a criminal. Meanwhile, researchers from Stanford demonstrate the problem of reproducibility by giving a data set of brain scans to 70 different researcher teams; no two teams chose the same workflow to analyze the data, and the final conclusions showed a sizeable variation. In a similar vein, researchers at Duke University examine the historical record of brain scan research and find poor correlation across experiments. In research, the "best paper" for the Conference on Computer Vision and Pattern Recognition goes to a team from Oxford, who use unsupervised learning methods and symmetry to convert single 2D images into 3D models. Researchers at Uber, the University of Toronto, and MIT use 3D simulated worlds to generate synthetic data for training LiDAR systems on self-driving vehicles. Calum MacKellar makes Cyborg Mind available, a look into the future of cyberneuroethics. And Johns Hopkins prepares for a second seminar on Operationalizing AI in Health.

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

For COVID-related AI news, Andy and Dave discuss the Stanford Social Innovation Review report on the problem with COVD-19 AI solutions (e.g., data gaps, inconsistency, etc), and how to fix them. The National Endowment for Science Technology and the Arts (NESTA) provides a thorough report on AI and COVID-19, whose findings generally suggest that barriers might exist for the employment of AI in tackling COVID-19. In regular AI news, the US has its first known case of an erroneous arrest due to facial recognition technology, with the arrest of Robert Williams in Detroit in January 2020 (and disclosed on 24 June). The European Commission white paper on AI gets two more responses, from Facebook and from the Center for Data Innovation. Sergei Ivanov provides a breakdown of contributors for the upcoming International Conference on Machine Learning. Researchers have identified a new threat vector against neural networks, one that increases energy consumption and latency. And a follow-up with the Pulse upsampling tool shows a bias toward producing white faces, likely inherited from its training dataset, StyleGAN. In research, Denny Britz examines replicability issues in AI research, and how academic incentive systems are driving the AI research community toward certain types of research. The Marine Corps University Journal turns into the Journal of Advanced Military Studies, and its first issue focuses on innovation and future war. The Combat Studies Institute Press publishes On Strategy: A Primer, including a chapter on future war by Mick Ryan. And Major Nicholas Narbutovskih pens Dust, a story about two warring factions with different approaches to autonomous systems.

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