Overparameterized Deep Neural Networks that generalize well have been key to the dramatic success of Deep Learning in recent years. The reasons for their remarkable ability to generalize are not well understood yet. It has also been known that deep networks possess the ability to memorize training data, as evidenced by perfect or high training accuracies on models trained with corrupted data that have class labels shuffled to varying degrees. Concomitantly, such models are known to generalize poorly, i.e. they suffer from poor test accuracies, due to which it is thought that the act of memorizing substantially degrades the ability to generalize. It has, however, been unclear why the poor generalization that accompanies such memorization, comes about. One possibility is that in the process of training with corrupted data, the layers of the network irretrievably reorganize their representations in a manner that makes generalization difficult. The other possibility is that the network retains significant ability to generalize, but the trained network somehow chooses to readout in a manner that is detrimental to generalization. Here, we provide evidence for the latter possibility by demonstrating, empirically, that such models possess information in their representations for substantially improved generalization, even in the face of memorization. Furthermore, such generalization abilities can be easily decoded from the internals of the trained model, and we build a technique to do so from the outputs of specific layers of the network. We demonstrate results on multiple models trained with a number of standard datasets.

In the supervised classification setting, during inference, deep networks typically make multiple predictions. For a pair of such predictions (that are in the top-k predictions), two distinct possibilities might occur. On the one hand, each of the two predictions might be primarily driven by two distinct sets of entities in the input. On the other hand, it is possible that there is a single entity or set of entities that is driving the prediction for both the classes in question. This latter case, in effect, corresponds to the network making two separate guesses about the identity of a single entity type. Clearly, both the guesses cannot be true, i.e. both the labels cannot be present in the input. Current techniques in interpretability research do not readily disambiguate these two cases, since they typically consider input attributions for one class label at a time. Here, we present a framework and method to do so, leveraging modern segmentation and input attribution techniques. Notably, our framework also provides a simple counterfactual "proof" of each case, which can be verified for the input on the model (i.e. without running the method again). We demonstrate that the method performs well for a number of samples from the ImageNet validation set and on multiple models.