Stabilize Deep ResNet with A Sharp Scaling Factor $τ$
This provides a theoretical foundation for stabilizing deep ResNet training, benefiting researchers and practitioners in deep learning by enabling more efficient and robust model training.
The paper tackles the stability and convergence of training deep ResNets by proving that scaling the residual branch by a factor τ = O(1/√L) ensures stable forward/backward processes and global convergence with gradient descent, independent of depth, and empirically shows this allows training without normalization layers and improves performance.
We study the stability and convergence of training deep ResNets with gradient descent. Specifically, we show that the parametric branch in the residual block should be scaled down by a factor $τ=O(1/\sqrt{L})$ to guarantee stable forward/backward process, where $L$ is the number of residual blocks. Moreover, we establish a converse result that the forward process is unbounded when $τ>L^{-\frac{1}{2}+c}$, for any positive constant $c$. The above two results together establish a sharp value of the scaling factor in determining the stability of deep ResNet. Based on the stability result, we further show that gradient descent finds the global minima if the ResNet is properly over-parameterized, which significantly improves over the previous work with a much larger range of $τ$ that admits global convergence. Moreover, we show that the convergence rate is independent of the depth, theoretically justifying the advantage of ResNet over vanilla feedforward network. Empirically, with such a factor $τ$, one can train deep ResNet without normalization layer. Moreover, for ResNets with normalization layer, adding such a factor $τ$ also stabilizes the training and obtains significant performance gain for deep ResNet.