2018/07/27 by Santiago A. Cadena, Cadena, Santiago A., Marissa A. Weis +7 · 1 citation
Computer Science · #Computational Physics and Python Applications #Neural Networks and Applications
paper · pdf · doi:10.48550/arxiv.1807.10589
Visualizing features in deep neural networks (DNNs) can help understanding\ntheir computations. Many previous studies aimed to visualize the selectivity of\nindividual units by finding meaningful images that maximize their activation.\nHowever, comparably little attention has been paid to visualizing to what image\ntransformations units in DNNs are invariant. Here we propose a method to\ndiscover invariances in the responses of hidden layer units of deep neural\nnetworks. Our approach is based on simultaneously searching for a batch of\nimages that strongly activate a unit while at the same time being as distinct\nfrom each other as possible. We find that even early convolutional layers in\nVGG-19 exhibit various forms of response invariance: near-perfect phase\ninvariance in some units and invariance to local diffeomorphic transformations\nin others. At the same time, we uncover representational differences with\nResNet-50 in its corresponding layers. We conclude that invariance\ntransformations are a major computational component learned by DNNs and we\nprovide a systematic method to study them.\n