BCD for DNN

This is the code for Jianqiu Lu's Undergraduate Thesis. It is used for training DNN by Block Coordinate Descent (Prox-Linear Update) in mini batch.

Results

We use MNIST dataset. Accuracy of Block Coordiante Descent training on 1 layer MLP, with structure of 784-800-10, is BCDM-1layerAccu while our baseline, backprop with SGD, has a accuracy of sgd_accu. Results on 3 layer MLP is BCDM-3layerAccu while baseline performance as Base_accu