Efficient coding of natural scenes improves neural system identification

Abstract

Neural system identification aims at learning the response function of neurons to arbitrary stimuli using experimentally recorded data, but typically does not leverage coding principles such as efficient coding of natural environments. Visual systems, however, have evolved to efficiently process input from the natural environment. Here, we present a normative network regularization for system identification models by incorporating, as a regularizer, the efficient coding hypothesis, which states that neural response properties of sensory representations are strongly shaped by the need to preserve most of the stimulus information with limited resources. Using this approach, we explored if a system identification model can be improved by sharing its convolutional filters with those of an autoencoder which aims to efficiently encode natural stimuli. To this end, we built a hybrid model to predict the responses of retinal neurons to noise stimuli. This approach did not only yield a higher performance than the stand-alone system identification model, it also produced more biologically-plausible filters. We found these results to be consistent for retinal responses to different stimuli and across model architectures. Moreover, our normatively regularized model performed particularly well in predicting responses of direction-of-motion sensitive retinal neurons. In summary, our results support the hypothesis that efficiently encoding environmental inputs can improve system identification models of early visual processing.

A chromatic feature detector in the retina signals visual context changes

“…Fig. S1a) (41). Cells assigned to any of groups 33-46 were considered displaced amacrine cells and were not analysed in this study (for detailed filtering pipeline, see Suppl.…”

“…The second and third step made sure only cells were included that were assigned to a ganglion cell group (i.e., group index between 1 and 32) with sufficient confidence. Confidence is defined as the probability assigned to the predicted class by the random forest classifier (see (41)), and the threshold was set at ≥ 0.25.…”

Nitric oxide modulates contrast suppression in a subset of mouse retinal ganglion cells

“…For each cell, Ca 2+ responses to chirp and moving bar, soma size, and p-value of permutation test for direction selectivity (left) constitute the input to the RFC (centre) to predict a cell type label, i.e., a type G X (right). For details, see Methods and (79). ( b ) Pooled heat map of unsorted cell responses to chirp and moving bar stimulus from both Ctrl 1 datasets after quality filtering (QI MB >0.6 or QI chirp >0.45, and classifier confidence score ≤ 0.25).…”

“…As we wanted to investigate if the tested drugs differentially affect the different retinal output channels, we applied an RGC type classifier (Fig. 2a) (80), which had been trained and validated on a previously published RGC Ca 2+ imaging dataset (5). The classifier predicts a GCL cell’s functional type based on soma size and the responses to chirp and moving bar stimuli (see Methods).…”

“…Classification of functional retinal ganglion cell types. For the functional classification of RGC types, we used a previously published RGC classifier (79). The classifier, which uses a random forest classifier, was trained, validated and tested on previously published RGC type responses (5).…”