Learning as statistical inference: neural and computational mechanisms for normative learning

• Date and time: October 3, 2017 from 11:00am-12:00 pm
• Location: BiocomplexityInstitute Conference Room
  
• Speaker:  Matthew Nassar, Ph.D.
  Postdoctoral Fellow
  Brown University
  

Successful decision-making often requires learning from prediction errors, but how much should we learn from any given error? I will examine this question in detail, drawing on an optimal inference model to formalize how we should learn in dynamic environments and a computationally efficient approximation to provide insight in to how we could do so by adjusting the rate of learning from moment to moment. I will show behavioral data validating key model predictions in humans, demonstrate a role for the arousal system in setting the learning rate, and dissect the computational roles of neural subsystems upstream of learning rate implementation. I will explore the possibility that learning deficits might emerge from a failure to correctly determine how much should be learned, rather than a failure to represent prediction errors per se, and provide evidence for such an explanation in the case of healthy aging. Finally I will re-examine neural architecture of error-driven learning in the context of these results and discuss some future directions emerging from this work.

Contact Anne Wailes for more information: awailes@vt.edu