Some developmental and attentional biases in the contrast enhancement and short term memory of recurrent neural networks

Abstract

This paper studies the global dynamics of neurons, or neuron populations, in a recurrent on-center off-surround anatomy undergoing nonlinear shunting interactions. In such an anatomy, a given population excites itself and inhibits other populations. The interactions are defined by multiplicative mass action laws. Grossberg (1973) studied the case in which all populations have the same weight (or total number of unit cell sites). Here the effect of an arbitrary distribution of population weights is studied; each set of populations with equal weight is called a subfield. Possible causes of variable population weights are developmental biases (e.g., which feature detectors are represented in a field), attentional changes (e.g., which features are relevant at any time), and statistical errors in network design. Such factors can bias the total field towards accentuating or suppressing in short-term memory a given subfield of sensory features. In particular, a mechanism is noted for suppressing the activity of populations whose trigger features are infrequently experienced by the network. These variables interact with the recurrent on-center off-surround interactions, that have previously been shown capable of contrast enhancing significant input information, sustaining this information in short-term memory, adapting the field's total activity while producing multistable equilibrium points of this activity, suppressing noise, and preventing saturation of population response even to input patterns whose intensities are high.