Explaining Distributed Neural Activations via Unsupervised Learning

Recent work has demonstrated the emergence of semantic object-part detectors in activation patterns of convolutional neural networks (CNNs), but did not account for the distributed multi-layer neural activations in such networks. In this work, we propose a novel method to extract distributed patterns of activations from a CNN and show that such patterns correspond to high-level visual attributes. We propose an unsupervised learning module that sits above a pre-trained CNN and learns distributed activation patterns of the network. We utilize elastic non-negative matrix factorization to analyze the responses of a pretrained CNN to an input image and extract salient image regions. The corresponding patterns of neural activations for the extracted salient regions are then clustered via unsupervised deep embedding for clustering (DEC) framework. We demonstrate that these distributed activations contain high-level image features that could be explicitly used for image classification.