Abstract
Background: It has been hypothesized that adaptation plays a role in efficient sensory coding. Consider the responses of a population of orientation-tuned neurons to the following stimulus ensembles: 1) all orientations occur equally often; 2) one orientation Θ is biased to occur with greater frequency. Absent adaptation, overrepresenting Θ would cause neurons tuned near Θ to respond more on average, and would produce greater response covariance among these neurons, reducing coding efficiency. In agreement with the efficient coding hypothesis, adaptation to a biased ensemble restores the response amplitudes and covariances evoked by an unbiased ensemble (Benucci et al., Nat. Neurosci., 2013). We propose a computational model of adaptation to explain these findings. Methods: The model consisted of a population of neurons with linear, orientation-selective receptive fields, divisive normalization, and anti-Hebbian learning of normalization weights. For each stimulus presentation, the divisive normalization pools are updated as follows: 1) Measure the products of neural responses for each pair of orientation-tuned neurons. 2) Increase the contribution of neuron i to the divisive normalization pool of neuron j in proportion to this product for pair ij, minus its long-term expected value (fixed for each neuron pair, determined by their relative tuning). We simulated the steady-state behavior and dynamics of this model in response to a biased ensemble of rapidly flashed gratings. Results: Adapting to the biased stimulus ensemble changed the normalization weights, such that the responses amplitudes and covariances were restored to values consistent with the unbiased stimulus ensemble. In agreement with neurophysiological and psychophysical measurements, the resulting tuning curves were suppressed near the overrepresented orientation Θ, and were shifted away from Θ towards flanking orientations. Conclusion: A model of adaptation, in which sensory neurons with greater covariance than is expected update to inhibit each other more, explains the efficient coding of biased stimulus ensembles.
Meeting abstract presented at VSS 2014