I just finished the work on the caffe-brewer, an attempt to get the most recent version of caffe building on Windows for my fertilized-forests library. Doing this, I had (again) to find replacements for some undefined gcc functions for VS2012 and decided to collect them here.
Visual Studio supports many ways of debugging. The immediate window comes in very handy when it comes to examining specific variable values (e.g. what's the value of the 20000th element of this vector?).
For the generation of test samples, it is sometimes necessary to only use truncated values, since they have to be in a specific range. Achieving this for values following a normal distribution is not straightforward to implement efficiently since the distribution should be left unchanged.
Recently I had to deal with probabilities for vectors in python. Surprisingly, handling them is not straightforward, for practical and numerical reasons. I want to quickly sum up my findings here.
I prepared a basic introductory python script for teaching. It shows how a perceptron can be created and how a basic feed-forward two-layer network can be used to learn to approximate the XOR function. I also included code to visualize the decision boundary of the ANNs during training.