ALS instead BP on mnist. The architecture is three layer neural network with relu or sigmoid activate function.
"dataset" file should be unzipped, should be put into the same profile with codes.
"dataset" just contain MNIST, for SVHN, should be downloaded for "test_32x32.mat" and "train_32x32.mat" from official web.
http://ufldl.stanford.edu/housenumbers/
For CIFAR, should download "cifar-10-matlab.tar.gz" from web.
https://www.cs.toronto.edu/~kriz/cifar.html
mc function refers https://www.mathworks.com/matlabcentral/fileexchange/26395-matrix-completion-via-thresholding several toolboxes should be downloaded and set first, refer to the procedure of "demo" from the link.
run "demo_als_matrix", "index" for adjusting activation function, "mc" for adjusting matrix completion function.