Initialization and Regularization of Factorized Neural Layers
This work addresses optimization challenges in deep learning for practitioners using factorized layers, offering incremental improvements in model compression, knowledge distillation, and attention mechanisms.
The paper tackles the problem of initializing and regularizing factorized neural layers, such as those used in model compression and multi-head attention, by proposing spectral initialization and Frobenius decay to improve performance. The result includes enabling low-rank methods to outperform unstructured sparsity and tensor methods in model compression, with specific gains like improved translation and pre-training performance.
Factorized layers--operations parameterized by products of two or more matrices--occur in a variety of deep learning contexts, including compressed model training, certain types of knowledge distillation, and multi-head self-attention architectures. We study how to initialize and regularize deep nets containing such layers, examining two simple, understudied schemes, spectral initialization and Frobenius decay, for improving their performance. The guiding insight is to design optimization routines for these networks that are as close as possible to that of their well-tuned, non-decomposed counterparts; we back this intuition with an analysis of how the initialization and regularization schemes impact training with gradient descent, drawing on modern attempts to understand the interplay of weight-decay and batch-normalization. Empirically, we highlight the benefits of spectral initialization and Frobenius decay across a variety of settings. In model compression, we show that they enable low-rank methods to significantly outperform both unstructured sparsity and tensor methods on the task of training low-memory residual networks; analogs of the schemes also improve the performance of tensor decomposition techniques. For knowledge distillation, Frobenius decay enables a simple, overcomplete baseline that yields a compact model from over-parameterized training without requiring retraining with or pruning a teacher network. Finally, we show how both schemes applied to multi-head attention lead to improved performance on both translation and unsupervised pre-training.