Probing the Structure and Functional Properties of the Dropout-induced Correlated Variability in Convolutional Neural Networks

Abstract

Convolutional neural networks (CNNs) have been used to model the biological visual system. Compared to other models, CNNs can better capture neural responses to natural stimuli. However, previous successes are limited to modeling mean responses; while another fundamental aspect of cortical activity, namely response variability, is ignored. How the CNN models capture neural variability properties remains unknown. Previous computational neuroscience studies showed that the response variability can have a functional role, and found that the correlation structure (especially noise correlation) influences the amount of information in the population code. However, CNN models are typically deterministic, so noise (and correlations) in CNN models have not been studied. In this study, we developed a CNN model of visual cortex that includes neural variability. The model includes Monte Carlo dropout, namely a random subset of units is silenced at each presentation of the input image, inducing variability in the model. We found that our model captured a wide-range of neural variability findings in electrophysiology experiments, including that response mean and variance scale together, noise correlations are small but positive on average, both evoked and spontaneous noise correlation are larger for neurons with similar tuning, and the noise covariance is low-dimensional. Further, we found that removing the correlation can boost trial-by-trial decoding performance in the CNN model.