Apr 12, 2019
Andy and Dave discuss Simulated Policy Learning (SimPLe), from Google Brain, which attempts to help reinforcement learning methods learn effective policies for complex tasks, such as Atari games (using the Atari Learning Environment, ALE); the method trains a policy in a simulated environment so that it achieves good performance in the original environment. From Google and Princeton University, the TossingBot learns to throw arbitrary objects into bins; research use “residual physics” to provide a baseline knowledge of the world (e.g., ballistics) to further improve tossing accuracies. Researchers at Rutgers demonstrate a probabilistic approach for reasoning the 3D shapes of unknown objects, as a robot manipulates its environment. DeepMind publishes results that use the AI itself to figure out where the AI will fail. And research from Northwestern, University of Chicago, and the Santa Fe Institute examines the dynamics of failure across science, startups, and security efforts. In clickbait-y news, scientists create an AI that can predict when a person will die (when in actual, they used machine learning methods to examine prediction of premature death, and compared with standard epidemiological approaches). Researchers create a memristor-based hybrid analog-digital computing platform to demonstrate deep-Q reinforcement learning. Microsoft demonstrates end-to-end automation of DNA data storage (21 hours to encode the word “hello”). The US Air Force is exploring AI-powered autonomous drones in its Skyborg program. Keen Security Lab of Tencent reports vulnerabilities of Telsa Autopilot, to include inducing the vehicle to switch lanes. A paper in the Springer AI Review Journal provides a survey of ML and DL frameworks and libraries for large-scale data mining. Los Alamos Labs publishes a survey of quantum algorithm implementations. Scott Cunningham publishes Causal Inference. Yaneer Bar-Yam makes a 2003 work, Dynamics of Complex Systems, available. Easley and Kleinberg publish Networks, Crowds, and Markets: Reasoning About a Highly Connected World. Andy highlights a sci-fi story from 2008 from Elizabeth Bear, Tideline. Paul Oh pens a fictional story of the Army’s C2 AI program, Project AlphaWare. The National Academies-Royal Society Public Symposium will hold a discussion on 24 May, AI: An International Dialogue. More videos appear from DARPA’s AI Colloquium. A website compiles datasets for machine learning. And Stephen Jordan provides a comprehensive catalog of quantum algorithms.