Does the neuronal noise in cortex help generalization?

One remarkable feature of neuronal activity in the mammalian cortex is the high level of variability in response to repeated stimuli. First, we used an open dataset, the Allen Brain Observatory, to quantify the distribution of responses to repeated presentations of natural movies. We find that even for their preferred moment in the movie clip, neurons have high variability which cannot be well captured by Gaussian or Poisson distributions. A large fraction of responses are better fit by log-normal or Gaussian mixture models with two components. These distributions are similar to activity distributions during training of deep neural networks using dropout. This poses the interesting hypothesis: is the role of cortical noise to help in generalization during learning?Second, to ensure the robustness of our results we analyzed electrophysiological recordings in the same areas of mouse visual cortex, again using repeated natural movie presentations and found similar response distributions. To make sure that the trial-by-trial variations we observe are not due exclusively to the result of changes in state, we constructed a population coupling model, where each neuron’s activity is coupled to a low-dimension version of the activity of all other simultaneously recorded neurons. The population coupling model can capture global, brain-wide activity fluctuations that are state-dependent. The residuals from this model also show non-Gaussian noise distributions.Third, we ask a more specific question: is the noise in the cortex more likely to move the representation of the stimulus in-class versus out-of-class? To address this question, we analyzed the responses of neurons across trials from multiple sections of different movie clips. We observe that the noise in the cortex better aligns to in-class variations. We argue that a useful noise for learning generalizations is to move from representations of different exemplars in-class, similar to cortical noise.