Скачать с ютуб MIT EI Seminar - Phillip Isola - Emergent Intelligence: getting more out of agents than you bake in в хорошем качестве

MIT EI Seminar - Phillip Isola - Emergent Intelligence: getting more out of agents than you bake in

MIT Embodied Intelligence Seminar - February 19, 2020 Abstract: One of the most satisfying things in AI research is to get more out than you put in. I will talk about three projects where we got something surprising out, beyond what was explicitly baked in. First, a simple inductive bias for unsupervised learning discovers “semantics”, despite that no labels are used during training. Second, physically plausible “videos” emerge in the latent space of a generative model, despite that the model was only fit to static images. Third capable and robust “creatures” self-assemble from a collection of primitive limbs, despite having no centralized controller. I will suggest that a goal in artificial intelligence is to get the most out while baking the least in.