Altering backward pass gradients improves convergence

In typical neural network training, the gradients in the backward pass is determined by the forward pass. As a result, the two stages are coupled. However, it is often seen that neural networks perform worse when gradients explode or decline. To address this, numerous approaches like Gradient Clipping (GC) and Adaptive Gradient Clipping (AGC) have been developed to enhance the gradient behaviour of networks without normalization layers during backward passes. These techniques decouple the backward and forward passes and modify the gradients adaptively. A possible drawback of clipping approaches is that they must be calculated for each weight tensor in each layer. We offer the PowerGrad Transform (PGT), a comparable approach that alters and enhances the gradient flow behaviour in the backward pass but is calculated only in the final softmax layer. It is very computationally efficient and outperforms both GC and AGC, resulting in improved performance in networks without batch normalization. PGT is easy to integrate into existing networks, requiring just a few lines of code, and significantly increases performance in non-BN ResNets. The impact is more pronounced on big datasets like as ImageNet, when networks do not fit all of the training data and there is some training headroom. PGT makes it possible for the network to better fit the training data while simultaneously improving its performance on the test set.