Robust learning of position invariant visual representations with OFF responses

Abstract

Event Abstract Back to Event Robust learning of position invariant visual representations with OFF responses Our ability to recognize objects in spite of variations in size or position is important for a successful interaction with the environment. One possible neural basis for this ability is the visual representations that have been found in the temporal lobe of primates. The firing rate of neurons in these regions have been shown to be object specific but largely invariant to the retinal position and scale of the object [1]. Recent experiments by DiCarlo and colleagues suggest that the position invariance of visual representations in inferotemporal cortex (IT) of monkeys are learned by exploiting the temporal contiguity of our environment [2,3]. The rational behind the learning paradigm DiCarlo tested is that the visual stimuli received before and after a saccade are likely to contain the same objects, but at different retinal positions. Therefore, position invariant representations can be learned by changing the sensory processing such that the firing rate of neurons in IT in response to the visual stimuli before and after saccade should be the same. Any learning rule that implements this idea has to compare the pre- and postsaccadic stimulus and therefore requires a memory trace that stores either the identity of the presaccadic stimulus or the firing rate of the neuron in response to that stimulus. Here, we propose that this memory trace could be represented by neuronal OFF responses. OFF responses are short periods of increased activity after the offset of a stimulus, which are displayed by many neurons in the visual system. Because these responses are stimulus specific while overlapping temporally with the initial response to the next stimulus, they introduce correlations between responses to stimuli that occur in temporal succession. Here, we show that a simple rate-based model that combines OFF responses with Hebbian learning and homeostatic plasticity allows (a) the unsupervised learning of position invariant representations, (b) the reproduction of the experimental results of Li and DiCarlo [2] and (c) the development of neuronal selectivity with a qualitative match to experimental results on neuronal selectivity in IT [4]. Our simulations show that learning with OFF responses implements a time scale invariance in that it is very robust to variations in the intersaccade interval. A comparison shows that the trace rule [5], the classical online rule for learning invariances, is more sensitive to fixation time variations. Our results suggest a new computational role for OFF responses in sensory learning and show that we may lose interesting learning mechanisms by concentrating on mean firing rates and neglecting the temporal response pattern of sensory neurons.