Minibatch Approximate Greatest Descent on CIFAR-10 Dataset

The proposed research incorporates control theory into numerical optimization to develop two-phase long and short term global minimum seeking strategy for weight learning in minibatch deep learning neural network (DLNN). First-order derivative methods to update weights using single-phase optimization method applied the first-order derivative methods to update weights using single-phase optimization method, which in turn disregards the initial point located at far away from the optimum solution. Therefore, the number of iteration and error rate is high due to the slow roll-off from the initial point to the solution. The implementation of SDAGD in Convolutional Neural Network (CNN) is tested on a larger image dataset - CIFAR-10. The preliminary result shows that SDAGD achieved a misclassification rate of 13.8% on CIFAR-10 dataset.