Practice makes (too) perfect: Hebbian learning and the persistence of overly trained behaviors in subcortical circuits

Summary

Date: 
September 24, 2019 - 12:00pm
Location: 
Bio Labs 1080
About the Speaker
Name: 
Sean Escola
Speaker Title: 
Assistant Professor
Speaker Affiliation: 
Columbia University

Over a century of work in experimental psychology and neuroscience has shown that the recall of memories and the performance of learned behaviors are determined by two principal variables: recency and practice. In this work, we develop a bottom-up and mechanistic understanding of the interplay between these variables in sequentially learned memories and behaviors. To do this, we begin by modeling sequential learning in a single neuron performing classification of random input patterns, and derive a mathematical expression for the neuron's forgetting curve, which quantifies the loss of old information as new information is learned. Because this simple model is unable to address the effects of practice during learning, however, we augment it with a second input pathway consisting of synaptic weights that are modified with associative Hebbian learning, leading to a generalized forgetting curve that additionally depends on the number of times that each pattern is repeated during training. In this model, patterns that are repeated multiple times during training become far more resistant to being overwritten, with near perfect recall long after patterns that are presented only once have been forgotten. We show that this is also true in a more elaborate neural network trained with reinforcement learning to perform a sequentially learned navigation task. Furthermore, due to the slow Hebbian learning in the second pathway, the signals from the two pathways gradually become aligned with one another through repeated practice, driving downstream units in similar ways. By this mechanism, control of the downstream population is gradually passed from a fast, flexible pathway with reward-based learning to a slow, robust pathway with associative learning. We suggest a neurobiological interpretation of this model, identifying the fast input with cortex, the slow input with thalamus, and the downstream population with striatum, the major locus of reinforcement learning in the brain. This interpretation provides a quantitative framework for understanding the formation of habits and the transfer of control from cortical to subcortical circuits as behaviors become automatized through extended practice.