Nov 16, 2018
Deep generative models can generate “spurious” samples (i.e. errors). Researchers from Université Paris-Saclay and PSL Research University explore a basic question, “Is it possible to get rid of all spurious samples [in deep generative models] without sacrificing coverage of a model?” Their research suggests a “Heisenberg Uncertainty”-like tradeoff between full coverage and spurious objects. DeepMind announces large-scale GAN training for natural image synthesis with high fidelity. And Andy discusses Topaz’s “AI Gigapixel,” an AI-driven software capability that intelligently adds information to photos to increase their resolution/size. In the paper of the work, researchers flip the Turing Test and ask humans what one word would they use to convince a human judge that they’re alive; the results are crappy. On a related note, Andy recalls Brian Christian’s achievement of being The Most Human Human. For books of the week, the UK’s Development, Concepts, and Doctrine Centre publishes the 6th edition of Global Strategic Trends; papers from the 3rd conference on the Philosophy and Theory of AI are available in a single publications; and Minsky’s Society of the Mind get a free hyperlinked online version (with the classic illustrations). In the video of the week, the Center for Technology Innovation asks “Who should answer the ethical questions surrounding AI?” And in the “silliness of the week,” a robot appears at a UK parliamentary meeting and “talks” to MPs about the future of AI in the classroom.