ZerO Initialization: Initializing Neural Networks with only Zeros and Ones
This addresses the initialization problem in deep learning, offering a deterministic alternative to random weights, though it appears incremental as it builds on existing network architectures.
The authors tackled the challenge of selecting appropriate initial variance for deep neural networks by introducing ZerO, a deterministic initialization scheme using only zeros and ones, which achieved state-of-the-art performance on datasets like ImageNet with ResNet.
Deep neural networks are usually initialized with random weights, with adequately selected initial variance to ensure stable signal propagation during training. However, selecting the appropriate variance becomes challenging especially as the number of layers grows. In this work, we replace random weight initialization with a fully deterministic initialization scheme, viz., ZerO, which initializes the weights of networks with only zeros and ones (up to a normalization factor), based on identity and Hadamard transforms. Through both theoretical and empirical studies, we demonstrate that ZerO is able to train networks without damaging their expressivity. Applying ZerO on ResNet achieves state-of-the-art performance on various datasets, including ImageNet, which suggests random weights may be unnecessary for network initialization. In addition, ZerO has many benefits, such as training ultra deep networks (without batch-normalization), exhibiting low-rank learning trajectories that result in low-rank and sparse solutions, and improving training reproducibility.