Towards an FPGA based reconfigurable computing environment for neural network implementations

Three computational characteristics can be attributed to neural networks: parallelism, modularity, and dynamic-adaptation. We argue that these computational characteristics of neural networks map nicely to fine-grained FPGA based reconfigurable computing architectures. Neural network architectures are decomposed into a set of parameterized neural computation modules and implemented in the FPGAs as hardware contexts. Control programs and tools are being created to support run-time instantiation of hardware contexts, and to assemble them into a neural network, as well as to manage the dynamic reconfiguration of the neural network modules. The forms of parallelism that can be exploited for neural network implementations on FPGA based reconfigurable computing environments are described.