High-performance computing for neuroinformatics using FPGA

The brain represents information through the ensemble firing of neurons. These neural processes are difficult to study in vivo because they are highly non-linear, dynamical and often time-varying. Hardware systems, such as the FPGA-based platforms, are very efficient in doing such studies given their intrinsic parallelism, reconfigurability and real-time processing capability. We have successfully used the Xilinx Virtex-6 FPGA devices to prototype the generalized Laguerre-Volterra model (GLVM), which is a rigorous and well-functioning mathematical abstraction for the description of neural processes from a system input/output relationship standpoint. The hardware system first conducts GLVM parameters estimation using the neural firing data from experiments; then it is able to predict the neural firing outputs based on the field estimated model coefficients and the novel model inputs. The hardware system has been prototyped and is proved very efficient in this study compared to our previous software model running on the Intel Core i7-2620M CPU (with Turbo Boost to 3.4 GHz). It achieves up to a 2,662 times speedup in doing GLVM parameters estimation and a 699 times speedup in conducting neural firing outputs prediction. The calculation results are very precise with the NMSE being successfully controlled at the 10^-11 scale compared to the software approach. This FPGA-based architecture is also significant to the future cognitive neural prostheses design.