Understanding and Leveraging the Learning Phases of Neural Networks
This work addresses a foundational issue in machine learning by providing insights into learning phases, which could impact training practices, though it is incremental in refining existing theories like the information bottleneck.
The paper tackles the problem of understanding neural network learning dynamics by analyzing reconstruction loss and prediction performance, empirically showing three distinct phases and proving their existence for single-layer networks, with a practical implication that pre-training for transfer learning should stop early, achieving up to 5% accuracy improvement in some cases.
The learning dynamics of deep neural networks are not well understood. The information bottleneck (IB) theory proclaimed separate fitting and compression phases. But they have since been heavily debated. We comprehensively analyze the learning dynamics by investigating a layer's reconstruction ability of the input and prediction performance based on the evolution of parameters during training. We empirically show the existence of three phases using common datasets and architectures such as ResNet and VGG: (i) near constant reconstruction loss, (ii) decrease, and (iii) increase. We also derive an empirically grounded data model and prove the existence of phases for single-layer networks. Technically, our approach leverages classical complexity analysis. It differs from IB by relying on measuring reconstruction loss rather than information theoretic measures to relate information of intermediate layers and inputs. Our work implies a new best practice for transfer learning: We show empirically that the pre-training of a classifier should stop well before its performance is optimal.