AI with AI

Andy and Dave discuss the latest in AI news, including an announcement from Facebook AI that it achieved state of the art computer vision performance with its SEER model, by learning from one billion (with a ‘b’) random, unlabeled, and uncurated public Instagram images, reaching 84% top-1 accuracy on 13k images from ImageNet. DARPA launches a new Perceptually-enabled Task Guidance (PTG) to help humans perform complex tasks (such as through augmented reality); the effort will include both fundamental research as well as integrated demonstrations. DARPA also announces research teams from its Semantic Forensics (SemaFor) effort at probing media manipulations. Chris Ume, a Belgian visual effects artist, releases four deepfake videos of Tom Cruise, using two NVIDIA GPUs, two months training time, and further days of processing and tweaking for each clip. Researchers at the University of Washington, Berkeley, and Google Research use the StyleGAN2 framework to create “time-travel photography,” which peels away the limitations of early cameras to reveal restored images of the original photos; the effort also involves the creation of a modern “sibling,” which then gets merged with the original. OpenAI publishes the discovery that neurons in its CLIP network respond to the same concept, whether literal, symbolic (e.g., a sketch), or conceptual (e.g., text); they also discover an absurdly simple attack, which involves places a stick with a word onto an item. The report of the week from UNICEF looks at Adolescent Perspectives on AI, with insights from 245 adolescents from five countries. Montreal.AI provides a 33-page “cheat sheet” with condensed information and links on AI topics. The book of the week from E-IR examines Remote Warfare: Interdisciplinary Perspectives. And the fun site of the week, MyHeritage, lets users animate photos, or “re-animate your dead loves ones.”