Abstract:
This thesis improves the training algorithm for Convolutional Neural Network by using self-adaptive learning rate. Tuning for appropriate the value of learning rate usually requires human intervention. In this work, the learning rate can be adaptive based on Taylors formula. This formula has the relationship among the root mean square errors changed, connection template, weights and biases changes. The proposed self-adaptive learning rate is calculated from the root mean square error and the error curve surface gradient. From the experimental results, this proposed system can reach the convergence with smaller number of training iterations than the traditional algorithm with a constant learning rate.