Predictive Coding and Cognitive Computation in Large-Scale Brain Circuits
Abstract
Project AbstractIn spite of stunning advances in artificial intelligence (AI), many present-day AI algorithms are conceptualized in terms of input-to-output transformation. A typical example is visual object recognition, where the input is an image and the output is a category. Even research on natural language processing largely deals with speech recognition or translation from one language toanother, both are essentially input-output mappings. By contrast, the brain does not passively receive and process external inputs. Our actions are guided by behavioral goals, this process depends on our brain~s ability to proactively infer or estimate external events, primarily through a hierarchy of top-down projections in the cerebral cortex. This point of view is broadly capturedby the idea of predictive coding, the underlying neural circuit mechanism remains unknown. Furthermore, existing AI systems are not multi-taskers and particularly poor at rule/contextdependent flexible decision-making.To delineate the neural mechanisms of proactive information processing and cognition in the brain, we will pursue our modeling of large-scale circuits of the primate brain (which was first developed thanks to the ONR support) focused on functional roles of top-down projections, decision-making system capable of multiple tasks, and novel learning mechanisms for adaptivechoice behavior including social cognition. The proposed research has three aims. In Aim 1, we will build a large-scale cortical circuit model of the primate, where feed forward and feedback projections are wired through a laminar structure and dynamically manifested in a frequency-dependent manner. This model will betested both with monkey and human data. Aim 2 will to devoted to extending our model from Aim 1 to a continuous network with neurons selective for stimulus feature such as spatial location or orientation of a visual stimulus. This will enable us to rigorously investigatepredictive coding in several ways, such as surround inhibition and mismatch negativity as physiological signatures of predictive coding. We will also examine distinct types of top-down projections, mediating prediction of sensory stimulation and selective attention respectively, in the cortex, as well as potential contributions by subcortical structures such as the thalamus and basal ganglia. Aim 3 is focused on synaptic plasticity and learning to perform multiple tasks and social cognition. We will train recurrent neural networks using reinforcement learning instead ofsupervised learning. In contrast to the tradition of using one network for a single task, we will train a single neural circuit for many tasks. Finally, we will incorporate belief learning and strategy into our modeling framework so that it is capable of social cognition. This will be donewith a focus on the prisoner~s dilemma task, which requires cooperation between players. Our modeling will benefit from a collaboration with Keren Haroush, who recently discovered single neuron activity from behaving monkeys that represents the other player~s intention to cooperate or defect in a prisoner~s dilemma task. Taken together, the proposed research program is notablefor its sweeping scope, its emphasis on detailed neural circuit mechanisms, and its potential new insights for future advances of genuinely cognitive and social AI.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jan 04, 2017
- Source ID
- N000141712041
Entities
People
- Xiao-Jing Wang
Organizations
- New York University
- Office of Naval Research
- United States Navy