Emergent patterns of task-specific neurons in deep neural networks

Abstract

Visual cognition has long been the subject of curiosity within the realm of deep learning. While much research has gone into the development of neural network models that can at times outperform humans, the underlying principles behind truly understanding visual concepts remain elusive. Utilizing a multitask learning paradigm, we first explore the capacity for networks to generalize to understand visual reasoning concepts. We introduce a simplified visual reasoning dataset to train several network architectures, including a recently proposed model built specifically for relational reasoning. We collect the best performing networks and view their behavior on a neuronal level: visualizing task selectivity through patterns of activations from each network layer. Finally, we adjust our focus to a simpler form of visual reasoning involving the extraction of single attributes from attribute compositions. Here, we are able to both visualize and quantify the neuron task selectivity that leads to generalization.