Paper on arxiv: http://arxiv.org/abs/1511.02583

Paper on ICML visualization workshop: https://www.dropbox.com/s/yb03ej5xte8d9gj/ICMLworkshop2016.pdf?dl=0

This is my implmentation of Batch Normalized Maxout NIN.

Original Matconvnet: https://github.com/vlfeat/matconvnet

This respository is my modification of Matconvnet.

You can install this modification as same as original installation: http://www.vlfeat.org/matconvnet/install/

vl_compilenn('enableGpu', true, 'cudaMethod', 'nvcc', ... 'cudaRoot', your-cuda-toolkit\CUDA\v6.5', ... 'enableCudnn', true, 'cudnnRoot', 'your-cudnn-root\cuda') ;

I used VS2013, CUDA-6.5 and cudnn-v4.

I added followwing functions:

Maxout units (GPU supported only)

Data augmentations (horizontal flipping / pad zeros and random cropping)

After installation, run "\example\cifar\cnn_cifar.m"

"unit1" is the number of maxout units

"piece1" is the number of maxout pieces

net.layers{end+1} = struct('type', 'conv', ... 'name', 'maxoutconv1', ... 'weights', {{single(orthonorm(1,1,unit0,unit1*piece1)), b*ones(1,unit1*piece1,'single')}}, ... 'stride', 1, ... 'learningRate', [.1 1], ... 'weightDecay', [1 0], ... 'pad', 0) ;

net.layers{end+1} = struct('type', 'bnorm', 'name', 'bn2', ... 'weights', {{ones(unit1*piece1, 1, 'single'), zeros(unit1*piece1, 1, 'single')}},'learningRate', [1 1 .5],'weightDecay', [0 0]) ;

net.layers{end+1} = struct('type', 'maxout','numunit',unit1,'numpiece',piece1) ;

add following to your net opts

-> net.meta.trainOpts.augmentation= true;