hide
Free keywords:
-
Abstract:
Sensory decision-making typically involves a choice between multiple alternatives and the evidence supporting each
alternative may comprise some number of independent signals (e.g., from different receptive fields). To make an optimal
choice, all signals supporting each alternative must be integrated before the results can be compared. In the brain, sensory
signals are integrated and represented by downstream neural responses that perform a non-linear transformation of sensory
input. Here we report that choice performance can be more accurate when it relies on integrated (averaged) neural
responses rather than on integrated (averaged) noisy sensory signals, at odds with the common-sense expectation that any
neural processing should degrade sensory information. The improvement is obtained when sensory signals s are drawn from
heavy-tailed (non-Gaussian) distributions P(s) and when non-linear neural response functions Ψ(s) broadly match δs ln P(s).
As proof-of-principle, we devised a decision-making toy-model in which a separate population of spiking neurons represented
evidence for each alternative. Specifically, we simulated responses of integrate-and-fire neurons to uncorrelated synaptic
inputs, with variance encoding an independent 'sensory signal'. In this toy framework, integrating neural responses rather
than sensory signals more than doubled choice performance, depending on assumptions. The intuitive reason for this
improvement is that compressive (saturating) responsiveness reduces the influence of outlier signals. Although not widely
known in neuroscience, this effect is exploited routinely in engineering and statistics by preprocessing samples with nonlinear
"influence functions" [e.g., Huber, Ronchetti (2009) Robust Statistics, John Wiley & Sons]. An intriguing question for
further work is whether neural response functions in vivo match or adapt to heavy-tailed input distributions.
