Overcoding-and-paring: a bufferless neural chunking model

Abstract

Event Abstract Back to Event Overcoding-and-paring: a bufferless neural chunking model Chunking is the process by which a single, purely spatial code, comes to represent a sequence of other purely spatial codes, the items. To assign unique chunk codes to sequences having common first items or prefixes, e.g., 'car' and 'cat', the brain must wait at least until a sequence's first distinguishing item is presented. However, for the chunk code to be associated with all sequence items via accepted neural learning principles, which require pre- and post-synaptic coactivity, the chunk code must activate during the first item and remain active throughout the sequence. One way to reconcile these opposing constraints is to store items in a first-in-first-out queue ''a short-term memory (STM) buffer''consisting of several disjoint, functionally-equivalent (apart from ordinal position represented) fields, or slots. This allows all items to be co-active, allowing the chunk code to be chosen after all items are present and thus be a function of the entire sequence. Most existing chunking models follow this approach [1-4]. However, there is no direct neurophysiological evidence for such an STM buffer, and there is mounting evidence against this view [5]. I therefore propose a radically different, bufferless, chunking/STM model, called, overcoding-and-paring (OP). Implicit in the following model description are two assumptions: i) chunks/items are represented by sparse distributed codes, and ii) cells at higher levels, e.g., in the chunking field, have a longer intrinsic activation duration, or, persistence, than lower levels cells. Suppose a two-item sequence, 'AB', is to be chunked. At t=1, item A is activated in an input field and gives gives rise to an oversized sparse distributed code, an overcode, in a chunking field. A is then bidirectionally and transiently associated with the overcode. When B presents at t=2: a) it causes some of the overcode's cells to deactivate, or, be pared out, leaving behind the final chunk code; b) the transient synaptic increases between A and the pared-out cells are depotentiated; c) the transient increases between A and the surviving cells are made permanent; and d) B is bidirectionally associated with the surviving cells. This leaves behind a chunk code that depends on the whole sequence and yet was activated on its first item, reconciling the opposing constraints without the need for an STM buffer. In essence, the progressive persistence property transfers the STM functionality needed for chunking from the item level to the chunking level, essentially turning the standard conception of STM on its head. Studies of neural activity during learning of new associations, reviewed in [6], are supportive of this scheme.