Experimentation scripts for MNIST

[accosmin]

To evaluate:

• regularization methods (none vs. L2 vs. variance vs. softmax vs. random inputs for classification)
• pooling layers (soft average, full, gauss)
• convolution connectivity factors (e.g. conn in {1, 2, 4, 8})
• the impact of adding multiple layers
• training methods (batch vs. stochastic)
• different activation functions

[accosmin]

Also generate a summary log file: [model] [training method] [criterion] [average test error]