6x6 pixel MNIST Classify

You can enjoy 6x6 pixel coarse-grained MNIST classification with more than 90% accuracy.

Please unzip the “6x6MNISTclassify” and find files to classify 6x6 pixel coarse-grained MNIST dataset, which I modified some of the files provided by Ruslan Salakhutdinov and Geoff Hinton (http://www.cs.toronto.edu/~hinton/MatlabForSciencePaper.html).
By executing main file “mnistclasify2LayerLoad6x6”, more than 90% accuracy is available even in this extremely low resolution with current setting. Please use these codes at your own risk. I hope your feedback on the modification and enjoy.

1.Points of the modification
1)The number of hidden layers is changed from three [784 500 500 2000 10] to two [784 500 2000 10]
2)In addition, also the number of units of layers are changed to be [36 36 36 10] for 6x6 pixel image classification instead of 28x28 pixel.

2.Summary of the files
Five files below 1)~5) are changed for the modification as follows:
1)“mnistclasify2LayerLoad6x6” (main file, changed from “mnistclasify”)
Line 20: Setting of pre-training maximum epoch, currently maxepoch=1
Line 21: Number of hidden layer and number of units are changed to be [36 36 36 10]
Line 40-49: Skipping 2nd hidden layer in original code
2)“converterLoad6x6” (changed from “converter”)
Line 48,50-51: Loading 6x6 pixel MNIST test data
Line 104-105: Loading 6x6 pixel MNIST train data
Line 27-28, 32-35, 58-61, 79-82, 87-90, 113-116, 127-128: Monitoring parameters
3)“backpropclassify2Layer” (changed from “backpropclassify”)
Line 17: Setting of post-training maximum epoch, currently maxepoch=50
Line 27-28: Monitoring parameters
Line 22, 34, 42, 65-67 : Changing for the hidden layer reduction
Line 91-93: Changing for the deleted middle hidden layer
Line 136-138, 142,146-150, 152-157, 167-168 : Changing for the hidden layer reduction
Line 107-110, 171-183: Displaying parameters and results
4)“CG_CLASSIFY_INIT2Layer” (changed from “CG_CLASSIFY_INIT”)
Line15-17, 20-21: Changing for the hidden layer reduction
5)“CG_CLASSIFY2Layer” (changed from “CG_CLASSIFY”).
Line15-22, 26-31, 39-41, 53-61, 65-66: Changing for the hidden layer reduction
6)“makebatches”, ”rbm” and “minimize” are not changed except skipping “fprint()” etc. .
7)6x6 pixel coarse-grained MNIST data* for train “MNIST_Train_Nox36.mat” and test ” MNIST_Test_Nox36.mat”, as well as original MNIST four data files of 28x28 pixel MNIST data are also included.
*https://jp.mathworks.com/matlabcentral/fileexchange/79580-coarse-grained-6x6-pixel-mnist-dataset