Initialization and Regularization of Factorized Neural Layers

By Mikhail Khodak, Neil A. Tenenholtz, Lester W. Mackey, Nicoló Fusi. 2021

78 citations recorded by the source corpus.

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.

Subfields: Computational Mathematics, Computer Vision and Pattern Recognition, Statistical and Nonlinear Physics, Optimization & Training

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computational Mathematics

  1. Tri Dao
  2. Christopher R'e
  3. Atri Rudra
  4. Moritz Hardt

In Computer Vision and Pattern Recognition

  1. Tri Dao
  2. Atri Rudra
  3. Christopher R'e
  4. Beidi Chen
  5. Nimit Sharad Sohoni
  6. Alexander Liu
  7. Michael Poli
  8. Alexander J. Ratner

In Statistical and Nonlinear Physics

  1. Christopher R'e

In Optimization & Training

  1. Tri Dao
  2. Christopher R'e
  3. Atri Rudra
  4. Nimit Sharad Sohoni
  5. Beidi Chen
  6. Peter Richtárik
  7. Nicolas Loizou
  8. Sebastian U. Stich

Citations in the corpus