Convolutional neural network The project is composed of two parts: 1.The code for forward and backpropagation through a network, and implement the necessary functions for training. 2.Design and train a CNN to classify numbers from the MNIST dataset.
Please open the train code for runnning.
The comparison of loss among different learning rates are shown in the following figures:
From the comparison among the development of loss function using different magnitude of LR, it can be proposed that using the LR about 0.05 can obtain about 96% accuracy after running 18000 iterations. With more iterations, the accuracy is expected to increase to some extent.