Published in Computational and Systems Neuroscience (CoSyNe), Feb 2019.
Accurate decision making requires reliable encoding and flexible decoding of sensory information. Given a specific task, only a small subset of neurons encode relevant information and need to be identified by a decoder depending on that task. No current theory explains how such decoding flexibility can be achieved, while respecting biological constraints on available knowledge of the encoding population. Haimerl & Simoncelli (Cosyne, 2018) introduced a framework for neural decoding based on functionally-targeted stochastic modulators, motivated by experimental observations in monkey V4 (Rabinowitz et al, 2015). That work revealed that despite its detrimental effect on encoding, the introduction of shared modulatory noise in task-informative neurons could serve as a labeling mechanism for decoding. Specifically, it introduced a modulator-guided decoding scheme which does not require biologically unrealistic knowledge assumed by an ideal observer. It only needs access to the low-dimensional modulator. Here we test predictions of this theory using population recordings from macaque V1 during a discrimination task (Ruff & Cohen, 2016). We first fit a Poisson Linear Dynamical System (Macke et al, 2015) to the data and extract two latent sources of co-variability, a slow drift and a fast modulator. We find that the fast modulator preferentially targets task-informative neurons as predicted by the theory, whereas the slow drift is task irrelevant. Second, we compare the performance of decoders that differ in their assumed knowledge about the encoding population and hence in their biological plausibility. We find that the modulator-guided decoder reaches the upper bound set by an ideal observer. Moreover, its parameters can be learned from a small number of training trials. These results confirm that targeted modulation could provide a mechanism enabling flexible and accurate decoding of neural responses in V1.