Normalization
is a widespread neural computation in both early sensory coding and
higher-order processes such as attention and multisensory
integration. It has been shown that during decision-making,
normalization implements a context-dependent value code in parietal
cortex. In this paper we develop a simple differential equations
model based on presumed neural circuitry that implements normalization
at equilibrium and predicts specific time-varying properties of value
coding. Moreover, we show that when parameters representing value
are changed, the solution curves change in a manner consistent with
normalization theory and experiment. We show that these dynamic
normalization models naturally implement a time-discounted
normalization over past activity, implying an intrinsic
reference-dependence in value coding of a kind seen
experimentally. These results suggest that a single network
mechanism can explain transient and sustained decision activity,
reference dependence through time discounting, and hence emphasizes the
importance of a dynamic rather than static view of divisive
normalization in neural coding.
|