Chunking space and time with information geometry

Humans are exposed to a continuous stream of sensory data, yet understand theworld in terms of discrete concepts. A large body of work has focused on chunkingsensory data in time, i.e. finding event boundaries, typically identified by modelprediction errors. Similarly, chucking sensory data in space is the problem at handwhen building spatial maps for navigation. In this work, we argue that a singlemechanism underlies both, which is building a hierarchical generative model ofperception and action, where chunks at a higher level are formed by segmentssurpassing a certain information distance at the level below. We demonstrate howthis can work in the case of robot navigation, and discuss how this could relate tohuman cognition in general.Humans are exposed to a continuous stream of sensory data, yet understand theworld in terms of discrete concepts. A large body of work has focused on chunkingsensory data in time, i.e. finding event boundaries, typically identified by modelprediction errors. Similarly, chucking sensory data in space is the problem at handwhen building spatial maps for navigation. In this work, we argue that a singlemechanism underlies both, which is building a hierarchical generative model ofperception and action, where chunks at a higher level are formed by segmentssurpassing a certain information distance at the level below. We demonstrate howthis can work in the case of robot navigation, and discuss how this could relate tohuman cognition in general.C