Stephen J. Guy

GPGU Class Project Proposal

Goal:

For my project I plan to implement Artificial Neural Networks (ANN) on a Graphics Processing Unit (GPU). ANNs are a very popular technique for approximating complex functions. It has been used for a verity of applications such as object/character recognition, stock market prediction, and automated driving.

I plan to implement a multi-layer ANN on the GPU. This most recent work in this area was from the China University of Geoscience in 2005 (http://ieeexplore.ieee.org/iel5/10421/33089/01555903.pdf?tp=&isnumber=&arnumber=1555903). This paper saw a good speed-up for their GPU implementation, but was week and describing the technique they used, and used an older generation of graphics hardware.

Novelty:

While programming ANNs on GPUs has been done before I still believe there is room for improvement. Furthermore, I hope to be able to implement training on the GPU, which (to my knowledge) has not been done before.

Milestones:

2/28 - Implement a simple ANN on the GPU

3/5 - Have explored limitations of implementation (Different Network structure, different threshold functions, etc)

3/16 - Implement a GPU based update function (Hopefully!)

4/1 - Add an interesting application using an ANN

Other Work on this topic:

• Kyoung-Su Oh, Keechul Jung, GPU implementation of neural networks, Patten Recognition 37, 6, 1311-1314, 2004

• Thomas Rolfes. Artificial Neural Networks on Programmable Graphics Hardware, in Game Programming Gems 4,pp373-378, 2004