Dynamically Partitionable Autoassociative Networks as a Solution to the Neural Binding Problem

摘要

An outstanding question in theoretical neuroscience is how the brain solves the neural binding problem. In vision, binding can be summarized as the ability to represent that certain properties belong to one object while other properties belong to a different object. I review the binding problem in visual and other domains, and review its simplest proposed solution – the anatomical binding hypothesis. This hypothesis has traditionally been rejected as a true solution because it seems to require a type of one-to-one wiring of neurons that would be impossible in a biological system (as opposed to an engineered system like a computer). I show that this requirement for one-to-one wiring can be loosened by carefully considering how the neural representation is actually put to use by the rest of the brain. This leads to a solution where a symbol is represented not as a particular pattern of neural activation but instead as a piece of a global stable attractor state. I introduce the Dynamically Partitionable AutoAssociative Network (DPAAN) as an implementation of this solution and show how DPANNs can be used in systems which perform perceptual binding and in systems that implement syntax-sensitive rules. Finally I show how the core parts of the cognitive architecture ACT-R can be neurally implemented using a DPAAN as ACT-R’s global workspace. Because the DPAAN solution to the binding problem requires only “flat” neural representations (as opposed to the phase encoded representation hypothesized in neural synchrony solutions) it is directly compatible with the most well developed neural models of learning, memory, and pattern recognition.