[PS-2.5] Error Driven Learning within the Hippocampus; theta rhythm, and novelty based learning signals

Ketz, N. , Gnanasekaran, S. & O'Reilly, R.

University of Colorado Boulder

Previous evidence has shown that the theta rhythm within the hippocampus is crucial for encoding and recall, but its mechanistic relation to learning and memory has yet to fully understood. Mathematical modeling work has suggested that different phases of this oscillatory signal facilitate hippocampal encoding or retrieval(Hasselmo, Bodelon & Wyble 2002, Hasselmo & Eichenbaum 2005). Similarly, neuroanatomical modeling has proposed the role of the subiculum, within a larger striatal circuit, as providing a dopamine facilitated novelty signal within area CA1(Lisman & Grace 2005). The current work expands upon these ideas by implementing and testing them within an existing, biologically plausible, neural network model of the hippocampus(Norman & O'Reilly 2003). Representations are first encoded during the theta trough in the auto-encoder like Entorhinal Cortex(EC) to CA1 connections. However, during the subsequent peak of the theta wave, retrieval occurs such that the CA3 drives a completed pattern of activation in CA1 where it is compared against the pattern encoded during the theta trough, providing an error-driven learning signal. Similarly, the subiculum provides a novelty signal based on the mismatch of incoming stimuli and the completed pattern of activation within EC, which in turn drives a dynamic learning rate in the CA3 to CA1 connections. Simulations show enhanced performance through the error-driven learning signal, as well as the dynamic learning rate. Performance is assessed in the AB-AC cued recall task, as well as a raw learning capacity test, where the augmented hippocampal model is compared with a purely Hebbian based hippocampal model. Results, explored across various network sizes, show a decrease in interference between studied items, as well as an increase in raw capacity for the augmented model compared to the Hebbian model. The current work is presented in the development of a mechanistic account of pre-frontal interactions with hippocampal encoding and retrieval.