Stefano Fusi, Columbia Neuroscience Center

The computational advantages of multi-stage memory systems

 

Over a century of experimental and clinical studies provide overwhelming evidence that declarative memory is a dynamical and spatially-distributed process.  Specifically, lesion studies have shown that the hippocampus is crucial for the formation of new memories, but that its role decreases in importance over time; ablation of the hippocampus does not affect remote memories.  This suggests that memory consolidation involves the transference of memory to extra-hippocampal areas, most likely the cerebral cortex.  Despite the wealth of behavioral data from animals and humans on this consolidation process, relatively little theoretical work has been done to understand it, and no work has addressed the underlying physiological process, which is presumably long-term synaptic plasticity.  The model consists of N plastic, binary synapses, divided into n stages.  Uncorrelated memories are encoded in the first stage with a rate r.  Synapses in the second stage are potentiated or depressed with a fixed probability according to the state (potentiated or depressed) of synapses in stage 1.  Synapses in downstream stages are updated in an analogous way with stage k directly influencing only stage k+1.  Additionally, synapses become increasingly less plastic the further downstream one goes, i.e. learning rates decrease with increasing stage number.  Therefore we posit a feed-forward structure in which the memory trace in each stage is actively transferred to the next downstream stage.  This is reminiscent of the physiological process of replay that has been recorded in cells of awake and sleeping rats.  It is also consistent with the clinical and experimental findings showing that memory consolidation appears to involve memory transfer to areas outside of the hippocampus.  The model trivially reproduces power-law forgetting curves for the learned memories by virtue of the distribution of learning rates.  Furthermore, through degradation or removal of early stages in our model we can easily account for the phenomena of anterograde and graded retrograde amnesia, which are common in animals and humans having suffered damage to the hippocampus.  In a similar vein we can qualitatively reproduce results from studies in which the administration of drugs has been found to selectively enhance or degrade memories in a temporally graded fashion.  Finally, this model leads to vastly improved memory traces compared to uncoupled synapses; this holds especially when adjacent stages have nearly the same learning rate and the total number of stages is large.  Joint work with Alex Roxin.