Mitigating the Vanishing Gradient Problem Using a Pseudo-Normalizing Method

When training a neural network, the choice of activation function can greatly impact its performance. A function with a larger derivative may cause the coefficients of the latter layers to deviate further from the calculated direction, making deep learning more difficult to train. However, an activation function with a derivative amplitude of less than one can result in the problem of a vanishing gradient. To overcome this drawback, we propose the application of pseudo-normalization to enlarge some gradients by dividing them by the root mean square. This amplification is performed every few layers to ensure that the amplitudes are larger than one, thus avoiding the condition of vanishing gradient and preventing gradient explosion. We successfully applied this approach to several deep learning networks with hyperbolic tangent activation for image classifications. To gain a deeper understanding of the algorithm, we employed interpretability techniques to examine the network’s prediction outcomes. We discovered that, in contrast to popular networks that learn picture characteristics, the networks primarily employ the contour information of images for categorization. This suggests that our technique can be utilized in addition to other widely used algorithms.