Learning in a Unitary Coherent Hippocampus

A previous paper [2] presented a model (UCPF-HC) of the hippocampus as a unitary coherent particle filter, which combines the classical hippocampal roles of associative memory and spatial navigation, using a Bayesian filter framework. The present paper extends this model to include online learning of connections to and from the CA3 region. Learning in the extended neural network is equivalent to learning in a temporal restricted Boltzmann machine under certain assumptions about neuromodulatory effects on connectivity and learning during theta cycles, which suggest detailed neural mappings for Bayesian inference and learning within sub-stages of a theta cycle. After-depolarisations (ADP) are hypothesised to play a novel role to enable reuse of recurrent prior information across sub-stages of theta.