Memory Modification in the Brain: Computational and Experimental Investigations

Abstract

I explore how and when memory traces are modified by new experience. Using a variety of paradigms, species and analytical tools, I argue that memories reflect inferences about the structure of the world. In particular, memories reflect the assignment of events to latent (hidden) causes. A new event modifies an existing memory trace if it is probable that the event was caused by the same latent cause as that represented by the old trace; otherwise, a new memory trace is formed. I show that probabilistic inference over latent causes, or \emph{structure learning}, provides a parsimonious explanation of many phenomena in human and animal learning, and may guide us towards developing new treatments for pathological memories like trauma and addiction. I first introduce a latent cause framework for modeling classical conditioning, based on ideas from modern Bayesian nonparametric statistics. Evidence suggests that an ostensibly extinguished memory can return under a variety of circumstances. The latent cause theory proposes that extinction training increases the probability that a new latent cause is active, thereby leading to the formation of two memories (one for acquisition, one for extinction). This theoretical explanation can also account for several other behavioral phenomena, as well as developmental trajectories and damage to the hippocampus. I argue that immature or hippocampally-damaged animals are impaired at expanding their repertoire of latent causes. I then develop a variant of the latent cause framework designed to explain the phenomenon of memory reconsolidation: retrieving a memory appears to render it temporarily labile. I show that the major phenomena of reconsolidation can be explained in terms of this framework, and I present new experimental data testing some of the theory's predictions. Motivated by this computational framework, I explore in several experiments the factors governing latent causal inferences by rats and humans. Taken together, these experimental and theoretical results support the idea that memory modification can be understood as a process of structure learning.