High performance Artificial Neural Network library.

Dependencies:

• FFTW3

As you may have guessed from the dependency on FFTW, this library uses FFT convolutions by default.

We use the CMake build system. Here is an example of how you might build the library:

mkdir build && cd build
cmake ..
make

Then to install, run this as root:

make install

If your CPU has access to AVX instructions you should use the -DENABLE_AVX=ON flag when running cmake.

Similarly, if your CPU has access to FMA instructions the -DENABLE_FMA=ON flag should be used as well.

Included is an example convolutional network for classifying MNIST. The network is based on one described in [1]. The program expects five filenames to be passed as command line arguments:

mnistcnn <train-images> <test-images> <train-labels> <test-labels> <output_model>

The first four files are the input MNIST data files from Yann LeCun's website [2]. The last filename is where the trained model is stored.

[1] Scherer, D., Müller, A., & Behnke, S. (2010). Evaluation of pooling operations in convolutional architectures for object recognition. In Artificial Neural Networks–ICANN 2010 (pp. 92-101). Springer Berlin Heidelberg.

[2] http://yann.lecun.com/exdb/mnist/

Please cite the following reference in papers using this library:

Gouk, H. G., & Blake, A. M. (2014, November). Fast Sliding Window Classification with Convolutional Neural Networks. In Proceedings of the 29th International Conference on Image and Vision Computing New Zealand (p. 114). ACM.