Predictive Coding Strategies for Invariant Object Recognition and Volitional Motion Control in Neuromorphic Agents
Abstract
Aim #1: Learning invariant representations of environments through experience has been important area of research both in the field of machine learning as well as in computational neuroscience. In this study, we employed a novel method for the discovery of invariants from a single video input based on the learning of the predictability of spatio-temporal relationships between inputs. To this end, videos containing spatio-temporal movements of unlabeled natural objects were used. Progress 1\202 Conducted real-time invariant perception and tracking of natural images 2\202 Conducted real-time invariant perception and tracking of video objects Aim #2: Volitional movements are a hallmark for human behavior. In this project, we hypothesized that visual memory of past motion trajectories may be used for selecting future behavior. In other words: following free energy principle, apparent volitional movements can be generated by minimizing the difference between what the agent expected to see and what it effectively sees. Progress 1\202 Tested of robotic systems prediction-based pseud-volitional movements in a complex environment requiring adaptive modifications. Phase I: physics modeled in Gazebo-type of simulated environment 2\202 Tested of robotic systems prediction-based pseud-volitional movements in a complex environment requiring adaptive modifications. Phase II: tested in a real physical environment
Document Details
- Document Type
- Technical Report
- Publication Date
- Sep 02, 2015
- Accession Number
- ADA626818
Entities
People
- Dae-shik Kim
Organizations
- KAIST