Published in Computational and Systems Neuroscience (CoSyNe), Mar 2018.
Sensory behaviors are often described in terms of encoding (from stimuli to neural responses) and decoding (from neural responses to behavior). Given full knowledge of encoding (stimulus selectivity, response variability), one can define statistically optimal decoders that provide an upper bound on performance. However, flexible and accurate decoding relies on identification of behaviorally-informative neurons (Britten et al., 1992). No current theory provides an explanation of how this can be achieved in the brain. As an example, in experiments by Cohen & Maunsell (2009) populations of V4 cells were monitored while a monkey performed an orientation discrimination task. A small subset of spatially scattered cells carried information relevant to the task in any given block of trials, and yet the animal performed the task well after viewing only a few examples. Optimizing even a linear decoder requires access to encoding details, or large amounts of labeled training data. A subsequent model-based analysis of these data (Rabinowitz et al., 2015) revealed that cells are gain-modulated by a common fluctuating low-dimensional signal, and neuron-specific strength of this modulation is correlated with task-specific informativeness. Here, we propose that this modulatory signal can serve as a key element in solving the identification step of decoding. We simulated a population of modulated Poisson neurons (Goris et al., 2014) responding to two stimuli with rate and modulation parameters matching those fit in Rabinowitz et al. (2015). We find that a linear decoder, with readout weights estimated from the strength of modulation, but without knowledge of the encoding model, achieves performance close to an optimal (ML) decoder assuming Poisson noise. Performance depends on the strength of the modulatory signal; poor at low and high levels and best for moderate levels. We conclude that fluctuating modulatory signals could be used to flexibly and rapidly identify task-specific neurons for decoding.