Coloring the Black Box: Visualizing neural network behavior with a self-introspective model

Abstract: The following work presents how autoencoding all the possible hidden activations of a network for a given problem can provide insight about its structure, behavior, and vulnerabilities. The method, termed self-introspection, can show that a trained model showcases similar activation patterns (albeit randomly distributed due to initialization) when shown data belonging to the same category, and classification errors occur in fringe areas where the activations are not as clearly defined, suggesting some form of random, slowly varying, implicit encoding occurring within deep networks, that can be observed with this representation. Additionally, obtaining a low-dimensional representation of all the activations allows for (1) real-time model evaluation in the context of a multiclass classification problem, (2) the rearrangement of all hidden layers by their relevance in obtaining a specific output, and (3) the obtainment of a framework where studying possible counter-measures to noise and adversarial attacks is possible. Self-introspection can show how damaged input data can modify the hidden activations, producing an erroneous response. A few illustrative are implemented for feedforward and convolutional models and the MNIST and CIFAR-10 datasets, showcasing its capabilities as a model evaluation framework.