Designing and training of the convolutional neural network

The structure of the designed convolutional neural network is shown in Fig. 5b. The CNN model consists of four convolutional layers, two max-pooling layers, and a fully connected layer, and the photocurrent matrix is programed to represent the convolution kernel to preprocess the image in the optoelectronic domain. A loss function was defined to quantify the performance of our model on images with known tags and was used to train the CNN. It is the percentage of inaccuracy between the predicted and observed markers. We employed a particular form known as Categorical for classification (> 2 classes). The RMSPROP update makes minor adjustments to the ADAGRAD method to lessen its aggressive and monotonically decreasing learning rate. Based on the data generated by the sensor array, a tailor-made data set consisting of eight types of motion (left to right, right to left, up to down, down to up, left-up to right-down, right-down to left-up, left-down to right-up and right up to left-down) was developed for training and testing. The learning rate (LR) annealing approach was utilized to make the optimizer converge more quickly and more closely to the global minimum of the loss function. The optimizer navigates across the loss landscape with LR. Larger steps and faster convergence are associated with greater LR. Nevertheless, the optimizer may encounter a local minimum due to the low sample quality and high learning rate. To effectively approach the global minimum of the loss function during training, it is preferable to have a decreasing learning rate. To maintain the benefit of a high learning rate in fast computation time, the learning rate will be halved if the accuracy does not increase after three epochs. After training 30 iterations in one epoch with a batch size of 86, the loss of train and test sets converged to a steady level, verifying that the CNN was well-trained and did not suffer from under- and over-fitting problems.