Organisms' actions and decisions are guided by experience. Models of such behavior often appeal to the formalism of probabilistic inference, in which expectations about the world build up sequentially due to past observations. These models can account for typical response patterns of subjects performing cog- nitive tasks. However, a theory grounded in biophysical principles of neural circuit architecture and activity is lacking. Our proposal seeks to ?ll this gap by constructing mechanistic neural circuit models of probabilistic infer- ence, which we will validate using innovative computational tools for matching the statistics of neural population recordings and subject behavior to the outputs of high-dimensional models. Our proposed work will address several outstanding questions concerning how neural circuits are guided by experience. Neural architecture likely plays a role in the brain's probabilistic computations, but there is not yet a clear theory of this connection. We propose that plasticity-driven changes in neural circuit architecture underlie these computations by reshaping the probability space of neural activity patterns. Neural activity is therefore biased to encode more likely beliefs, in light of experience. Our framework demonstrates this clearly using innovative mathematical methods to extract the low-dimensional activity dynamics of neural circuits subject to plasticity with various timescales. This approach will be applied to interpret our collaborators' data from subjects performing tasks in which they must estimate a remembered variable after a time delay. Theory is also lacking concerning how dynamics of neural activity represent variables that relevant to a cog- nitive tasks spanning multiple timescales. Most studies consider cleanly structured networks or purely random networks, producing fairly stereotypical neural population activity patterns. We will test the computational capa- bilities of plastic networks with mixed structured and random connectivity, focusing on how the resulting neural population dynamics represent remembered variables. Our neural circuit models will be validated and parame- terized using statistics of (a) neural populations recorded using multielectrodes in non-human primates and (b) subjects' behavioral responses. Our neural circuit models, software and tools used for ?tting our models, and data used to validate will be shared widely as a tool kit for use by the broader research community.
Humans have a remarkable capacity for adapting to and learning from experience, but these abilities can be disrupted by neurological disorders, aging, and brain trauma. To determine how best to treat pathologies that damage neural systems underlying learning, it is important to develop physiologically-grounded neural circuit models of how these systems function. The proposed research will develop a theory and tool kit for understanding how neural circuit architecture shapes such behavior using a combination of large-scale modeling, mathematical and computational analysis, and statistical analysis for neural population data.
