I am doing a thesis project. I need to make a FPGA implementation of a neural network. I want to implement some mathematical functions such as an exponential function.
Does anyone know of any good books about FPGAs, VHDL or digital architecture implementation of floating point arithmetic operations?
There are ANN techniques that don't require floating-point. You will find things much easier if you adopt one of those.
Jim Austin's work at York University comes to mind. He uses FPGAs and fixed-point DSPs.
EDIT: (Leon, I hope you don't mind me adding this list here)
"fpga neural networks" discussion forum, in particular a FPGA NN where "The neuron uses 8bit numbers".
"FPGA Implementations of Neural Networks - a Survey of a Decade of Progress" by Zhu and Sutton: lists many implementation techniques, and briefly mentions that floating point implementation technique has not yet been successful.
"FPGA neural net" search on Google Scholar
Professor Miriam Leeser's research group at the Reconfigurable Computing Laboratory at Northeastern University has developed an open source floating point library for Xilinx FPGAs in VHDL: http://www.ece.neu.edu/groups/rcl/projects/floatingpoint/index.html
I would avoid IEEE floating point format, go with the format the TI DSP uses. Ti has some good doc that goes along with one of the older DSPs that holds your hand through the operations. From there you can start on your logic.
