Is Robustness To Transformations Driven by Invariant Neural Representations

Deep Convolutional Neural Networks (DCNNs) have demonstrated impressive robustness to recognize objects under transformations (e.g. blur or noise) when these transformations are included in the training set. A hypothesis to explain such robustness is that DCNNs develop invariant neural representations that remain unaltered when the image is transformed. Yet, to what extent this hypothesis holds true is an outstanding question, as including transformations in the training set could lead to properties different from invariance, e.g. parts of the network could be specialized to recognize either transformed or non-transformed images. In this paper, we analyze the conditions under which invariance emerges. To do so, we leverage that invariant representations facilitate robustness to transformations for object categories that are not seen transformed during training. Our results with state-of-the-art DCNNs indicate that invariant representations strengthen as the number of transformed categories in the training set is increased. This is much more prominent with local transformations such as blurring and high-pass filtering, compared to geometric transformations such as rotation and thinning, that entail changes in the spatial arrangement of the object. Our results contribute to a better understanding of invariant representations in deep learning, and the conditions under which invariance spontaneously emerges.