Numerically efficient and biophysically accurate neuroprocessing platform

This paper presents a neuroprocessing system based on floating point arithmetic and a multi-core architecture in which optimized neuroprocessor cores model with biophysical accuracy different neuron sections like soma, dendrite and synapse. The system is focused on extracting detail information on the ion-channel dynamics and membrane voltage changes in single neurons (or groups of them) rather than implementing large neural networks; this details information is important from a neuroscience point of view. The neuroprocessors use numerical methods and floating point accuracy to solve the differential equations to create neuron representations based on the biological-compatible Hodking-Huxley and Traub models. The advanced extensible interface (AXI) interconnects the neuroprocessors to a central programmable processor in charge of monitoring, parameter loading and data distribution. The exponential operations involved in the modeling of the membrane voltage are optimized with floating-point look-up-tables. This approach reduces the computational time by 70% and complexity by around 40%. The accuracy and computation capabilities of the system are validated with experiments that involve the detection and discrimination of temporal input sequences, which is a fundamental brain function that underlies perception, cognition and motor output. Finally, a complete FPGA-PC platform is developed, the FPGA-based system interacts with a software interface in order to configure and receive results from the system running in hardware.