Climbing the Ladder of Causation Toward Mid-Level Vision

Explore a 58-minute lecture on mid-level vision and causal reasoning in artificial intelligence presented by Dan Yamins from Stanford University at IPAM's Analyzing High-dimensional Traces of Intelligent Behavior Workshop. Delve into the origins of rich low-to-mid-level visual scene understanding concepts in humans and animals, including contour estimation, optical flow, monocular depth perception, and 3D shape recognition. Discover a working theory based on Counterfactual World Models (CWMs) that explains how these concepts arise without detailed supervision. Learn about a specific form of masked prediction that enables training large-scale predictive models with causally-informative tokens. Examine how various mid-level visual concepts emerge through simple generic interventions on these tokens and the computation of counterfactual effects. Gain insights into a potential unsupervised algorithm for visual scene understanding in machines and consider novel hypotheses for the origins of biological vision.