AI with AI
Episode 2.19: A Neural Reading rAInbow
Andy and Dave discuss research from Neil Johnson, who looked to the movements of fly larvae to model financial systems, where a collection of agents share a common goal, but have no way to communicate and coordinate their activities (a memory of five past events ends up being the ideal balance). Researchers at Carnegie Mellon demonstrate that random search with early-stopping is a competitive Neural Architecture Search baseline, performing at least as well as “Efficient” NAS. Unrelated research, but near-simultaneously published, from AI Lab Swisscom, shows that random search outperforms state-of-the-art NAS algorithms. Researchers at DeepMind investigate the possibility of creating an agent that can discover its world, and introduce NDIGO (Neural Differential Information Gain Optimization), designed to be “information seeking.” And the Electronics and Telecomm Research Institute in South Korea creates SC-FEGAN, a face-editing GAN that builds off of a user’s sketches and other information. Georgetown University announces a $55M grant to create the Center for Security and Emerging Technology (CSET). Microsoft workers call on the company to cancel its military contract with the U.S. Army. DeepMind uses machine learning to predict wind turbine energy production. Australia’s Defence Department invests ~$5M to study how to make autonomous weapons behave ethically. And the U.K. government invests in its people and funds AI university courses with £115. Reports suggest that U.S. police departments are using biased data to train crime-predicting algorithms. A thesis on Neural Reading Comprehension and Beyond by Danqi Chen becomes highly read. A report looks at the evaluation of citation graphs in AI research, and researchers provide a survey of deep learning for image super-resolution. Bryon Reese blogs that we need new words to adjust to AI (to which Dave adds “AI-chemy” to the list). In Point and Counterpoint, David Sliver argues that AlphaZero exhibits the “essence of creativity,” while Sean Dorrance Kelly argues that AI can’t be an artist. Interpretable Machine Learning by Christoph Molnar hits version 1.0, and Andy highlights Asimov’s classic short story, The Machine that Won the War. And finally, a symposium at Princeton University’s Institute for Advanced Studies examines deep learning – alchemy or science?