THIS GRANT IS A CONTINUATION OF N000141410670 Bidirectional Vision
Abstract
The performer will conduct research to advance understanding of the bidirectional constrating satisfaction processing within the object recognition system and between it and spatial attention supported by the dosal pathway. The work will build upon the performer s existing object recognition model that uses bidirectional bottom-up and top-down connections during learning and recognition, and demonstrates some of the functional advantages of these bidirectional connections. The effort will tightly coordinate computational modeling and experimental research, using cutting-edge methods in humans (including fMRI, ERP, pharmacological manipulations, and careful experimental design) and macaque monkeys (including electrophysiological recordings, optogenetic manipulations of top-down connections, and pharmacology). Progress: Expenditures are reported at 90% in STARS. The performer will be out of funding by 31Dec2015. In FY15 the performer advanced our computational models in a number of important ways, and the experimental tests of these models are under way, using coordinated ERP and macaque electrophysiology experiments. For the models, they have developed a major new framework for understanding how the deep layers of the cortex (layers 5 and 6), interacting with the thalamus, provide a multiplicative top-down attentional mask that enables the network to focus specificially on the figure object while tuning down inputs from the background. These deep layers also support a local temporal context signal that enables powerful predictive learning over time, and a local auto-encoder learning mechanism to support local error-driven learning. All three of these mechanisms are synergistic and leverage the same neural circuits in multiple ways, and our simulations show that they lead to significant improvements in object recognition in cluttered visual scenes. They refer to this new neural framework as DeepLeabra, building on our existing model of superficial cortical layer dynamics (layers 2-3), called Leabra, that includes bidirectional attractor dynamics. The deep layer network updates at the alpha frequency (10 Hz, every 100 msec), and this provides a temporal envelope to coordinate learning and processing across the deep and superficial layers, and between cortex and thalamus. These time dynamics correspond to an expectation-maximization (EM) algorithm at a computational level, and explain an extensive amount of data on the alpha frequency properties of sensory cortex, including evidence of discrete elements of perception at the alpha frequency. Thus, the new DeepLeabra model provides a major new theoretical framework for integrating a very wide range of existing biological and behavioral data. They are in the process of writing up this model in the context of dorsal -- ventral visual pathway interactions to support a wide range of overt and covert visual attentional phenomena.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jun 03, 2016
- Source ID
- N000141612128
Entities
People
- Randall O Reilly
Organizations
- Office of Naval Research
- Regents of the University of Colorado
- United States Navy