Title: DeepTune: Visualization and interpretation of deep-network-based models in neuroscience
Authors: Yuansi Chen - ETH Zurich (Switzerland) [presenting]
Abstract: Deep neural network models have been shown effective recently in predicting single neuron responses in primate visual cortex area V4. V4 is a midtier visual cortical area in the ventral visual pathway. Its functional role is not yet well understood. Despite the high predictive performance of deep neural network models, these models are generally difficult to interpret. This limits the applicability of these models in characterizing V4 neuron function. We propose the DeepTune framework as a way to elicit interpretations of deep neural network-based models of single neurons in area V4. Using a dataset of recordings of 71 V4 neurons stimulated with thousands of static natural images, we built an ensemble of 18 neural network-based models per neuron to accurately predict its response given a stimulus image. To interpret and visualize these models, we used a stability criterion to form optimal stimuli (DeepTune images) by pooling the 18 models together. These DeepTune images not only confirm previous findings on the presence of diverse shape and texture tuning in area V4, but also provide concrete and naturalistic visualization of predicted optimal stimuli of individual V4 neurons.