Efficiency and Redundancy in Deep Learning Models : Theoretical Considerations and Practical Applications

Deep Neural Networks led to major breakthroughs in artificial intelligence. This unreasonable effectiveness is explained in part by a scaling-up in terms of computing power, available datasets and model size -- the latter was achieved by building deeper and deeper networks. In this thesis, recognizing that such models are hard to comprehend and to train, we study the set of neural networks under the prism of their functional equivalence classes in order to group networks by orbits and to only manipulate one carefully selected representant. Based on these theoretical considerations, we propose a variant of the stochastic gradient descent (SGD) algorithm which amounts to inserting, between the SGD iterations, additional steps allowing us to select the representant of the current equivalence class that minimizes a certain energy. The redundancy of the network's parameters highlighted in the first part naturally leads to the question of the efficiency of such networks, hence to the question of their compression. We develop a novel method, iPQ, relying on vector quantization that drastically reduces the size of a network while preserving its accuracy. When combining iPQ with a new pre-conditioning technique called Quant-Noise that injects quantization noise in the network before its compression, we obtain state-of-the-art tradeoffs in terms of size/accuracy. Finally, willing to confront such algorithms to product constraints, we propose an application allowing anyone to make an ultra-low bandwidth video call that is deployed on-device and runs in real time.