Simon Carbonnelle

Simon Carbonnelle

The prior knowledge (a.k.a. priors) integrated into the design of a machine learning system strongly influences its generalization abilities. In the specific context of deep learning, some of these priors are poorly understood as they implicitly emerge from the successful heuristics and tentative approximations of biological brains involved in deep learning design. Through the lens of supervised image classification problems, this thesis investigates the implicit integration of a natural clustering prior composed of three statements: (i) natural images exhibit a rich clustered structure, (ii) image classes are composed of multiple clusters and (iii) each cluster contains examples from a single class. The decomposition of classes into multiple clusters implies that supervised deep learning systems could benefit from unsupervised clustering to define appropriate decision boundaries. Hence, this thesis attempts to identify implicit clustering abilities, mechanisms and hyperparameters in deep learning systems and evaluate their relevance for explaining the generalization abilities of these systems. We do so through an extensive empirical study of the training dynamics as well as the neuron- and layer-level representations of deep neural networks. The resulting collection of experiments provides preliminary evidence for the relevance of the natural clustering prior for understanding deep learning.

Simon Carbonnelle, Christophe De Vleeschouwer

Our work presents extensive empirical evidence that layer rotation, i.e. the evolution across training of the cosine distance between each layer's weight vector and its initialization, constitutes an impressively consistent indicator of generalization performance. In particular, larger cosine distances between final and initial weights of each layer consistently translate into better generalization performance of the final model. Interestingly, this relation admits a network independent optimum: training procedures during which all layers' weights reach a cosine distance of 1 from their initialization consistently outperform other configurations -by up to 30% test accuracy. Moreover, we show that layer rotations are easily monitored and controlled (helpful for hyperparameter tuning) and potentially provide a unified framework to explain the impact of learning rate tuning, weight decay, learning rate warmups and adaptive gradient methods on generalization and training speed. In an attempt to explain the surprising properties of layer rotation, we show on a 1-layer MLP trained on MNIST that layer rotation correlates with the degree to which features of intermediate layers have been trained.