Studying Invariances of Trained Convolutional Neural Networks
This addresses the limited theoretical understanding of invariance learning in CNNs for image classification, though it is incremental as it builds on existing architectures.
The paper investigated how trained CNNs like AlexNet and ResNet learn invariances to image transformations, finding they show invariant behavior up to class- and transformation-dependent thresholds. It introduced the Invariant Transformer Net, a learnable module to extract the space of transformations to which CNNs are invariant and robust in class predictions.
Convolutional Neural Networks (CNNs) define an exceptionally powerful class of models for image classification, but the theoretical background and the understanding of how invariances to certain transformations are learned is limited. In a large scale screening with images modified by different affine and nonaffine transformations of varying magnitude, we analyzed the behavior of the CNN architectures AlexNet and ResNet. If the magnitude of different transformations does not exceed a class- and transformation dependent threshold, both architectures show invariant behavior. In this work we furthermore introduce a new learnable module, the Invariant Transformer Net, which enables us to learn differentiable parameters for a set of affine transformations. This allows us to extract the space of transformations to which the CNN is invariant and its class prediction robust.