A Workflow Foundation based Machine Learning Algorithm Library by using GPGPU for computation backend.
http://neuroflowblog.wordpress.com/
v0.0.2 Description
This is a very basic proof-of-concept implementation of the proposed library. There is a tiny example program for performance testing, and various unit tests for showing and verifying the features.
v0.0.2 Changelog
- Improved kernel compilation time requirement (binary caching implemented)
- Minor performance optimizations
v0.0.2 Features:
- It supports OpenCL CPU and GPU modes, kernels are optimized for each
- Architectures:
- Feed Forward Multilayer Perceptron
- Recurrent Multilayer Perceptron w/ Backpropagation Through Time
- Recurrent Multilayer Perceptron w/ Realtime Recurrent Learning (no OpenCL support yet)
- Learning algorithms:
- Online Gradient Descent
- Offline Gradient Descent
- Alopex-B (no OpenCL support yet)
- Cross Entropy Method (no OpenCL support yet)
v0.0.2 Requirements
- Visual Studio 2013
- Boost 1.55b (BOOST, BOOSTLIBx86, BOOSTLIBx64 environment variable must be set)
- Intel or AMD OpenCL SDK (OCLINC, OCLLIBx86, OCLLIBx64 environment variable must be set)
v0.0.2 Future Directions
- RTLR by using OpenCL CPU and GPU modes
- Multiple activation function support (Tanh, Bipolar Logistic)
- Cross Entropy cost function support
- Sign Changes, SuperSAB, Rprop (all variations), SCG, Oja
- Weight Decay
- Long Short Term Memory architecture
- Workflow Foundation 4.5 Activities. I have proof-of-concept code for WF integration ideas in my private repository but there is nothing publish-ready yet.
- Double precision support