Analyzing composability in a sparse encoding model of memorization and association

Abstract

A key question in neuroscience is how memorization and association are supported by the mammalian cortex. One possible model, proposed by Valiant, uses sparse encodings in a sparse random graph, but the composability of operations in this model (e.g. an association triggering another association) has not previously been evaluated. We evaluate composability by measuring the size of ldquoitemsrdquo produced by memorization and the propagation of signals through the ldquocircuitsrdquo created by memorization and association. While the association operation is sound, the memorization operation produces ldquoitemsrdquo with unstable size and produces circuits that are extremely sensitive to noise. We therefore amend the model, introducing an association stage into memorization. The amended model preserves and strengthens the sparse encoding hypothesis and invites further characterization of properties such as capacity and interference.