Abstract
Adaptive behavior in humans, rodents, and other animals often requires the integration over time of multiple sensory inputs. Here we studied the behavior and the neural activity of mice trained to actively integrate information from different whiskers to report the curvature of an object. The analysis of high speed videos of the whiskers revealed that the task could be solved by integrating linearly the whisker contacts on the object. However, recordings from the mouse barrel cortex revealed that the neural representations are high dimensional as the inputs from multiple whiskers are mixed non-linearly to produce the observed neural activity. The observed representation enables the animal to perform a broad class of significantly more complex tasks, with minimal disruption of the ability to generalize to novel situations in simpler tasks. Simulated recurrent neural networks trained to perform similar tasks reproduced both the behavioral and neuronal experimental observations. Our work suggests that the somatosensory cortex operates in a regime that represents an efficient compromise between generalization, which typically requires pure and linear mixed selectivity representations, and the ability to perform complex discrimination tasks, which is granted by non-linear mixed representations.
Competing Interest Statement
The authors have declared no competing interest.
