Incorporating neuronal fatigue in deep neural networks captures dynamics of adaptation in neurophysiology and perception

Abstract

The ability to rapidly adapt to the current sensory environment is a fundamental property of sensory systems. Such adaptation may depend on computations implemented at the neural circuit level, but also on intrinsic cellular mechanisms. We combined neural data, psychophysics, and computational models to evaluate whether an intrinsic neural fatigue mechanism in a feedforward hierarchical network is sufficient to explain the core properties of visual adaptation. We implemented activity-based response suppression in each unit of a convolutional deep neural network. The resulting units showed hallmark properties of repetition suppression, leading to a stimulus-specific and layerdependent response attenuation for frequent input. Furthermore, the activation patterns could account for known perceptual aftereffects such as the tilt and face-gender aftereffect. The model was also able to capture the results of a visual categorisation experiment demonstrating enhanced object recognition when objects are embedded in a constant pattern of noisy inputs. The computations required for adaptation can be learned, as demonstrated by units in a recurrent neural network which learn to suppress themselves when trained on that same psychophysics visual recognition experiment. These results show that a learned intrinsic neural mechanism of fatigue can incorporate temporal context information accounting for key neurophysiological and perceptual properties and leading to efficient and enhanced processing of sensory inputs. adaptation | object recognition | deep neural network | ventral stream | perceptual aftereffects | repetition suppression | neurophysiology | neuronal fatigueCorrespondence: kasper.vinken@childrens.harvard.edu