Abstract
We propose a theoretical framework for efficient representation of time-varying
sensory information using dynamic population codes. Our approach is based on
the hypothesis that for accurate perception and computation, it must be possible
for downstream neurons to readily extract correct estimates of stimulus states from the sensory input. Intuitively, optimal computational decoding must recover most of the encoded information. However, we have recently shown that even in a fairly constrained and analytically tractable formulation of a dynamic setting, decoding correct estimates can be a difficult computational problem. Information carried by the spikes is only temporally relevant, and when the input spikes are noisy or sparse, it becomes necessary to maintain a spiking history to perform accurate inference at any given time. We posit a recurrently connected population of neurons that recodes the input representation such that each spike can be decoded independently in a causal manner, without referring to any spike history. Decoding is carried out by a computationally simple,
biologically reasonable scheme that interprets spiking activity as representing a
probability distribution over stimulus stat
es. Coding then involves learning to
generate an apposite representation that
optimizes the fidelity of decoding. We
evaluate the efficacy of the proposed
coding scheme by assessing the capability
of the simple decoder in extracting the available information.
