Neuroinformatics and neuroprosthetics based on high-performance reconfigurable hardware platforms

基於高性能可重構硬件平台的神經信息學和神經義肢學

Student thesis: Doctoral Thesis

View graph of relations

Author(s)

  • Xiangyu LI

Related Research Unit(s)

Detail(s)

Awarding Institution
Supervisors/Advisors
Award date14 Feb 2014

Abstract

Cognitive neural prostheses are integrated electronic devices which aim at restoring the brain modalities damaged by injury or disease. Such prosthetic devices are expected to perform bidirectional communications between intact brain regions, bypassing the damaged region. These prostheses, if successfully developed, would provide fundamental treatment to diseases related to cognitive impairment such as the Alzheimer's disease. In this dissertation, I introduce our work regarding customizable and efficient architectural design for early-stage prototyping of the prosthesis using the Field-Programmable Gate Array (FPGA). The mathematical model implemented by our reconfigurable hardware platform is the generalized Laguerre-Volterra model (GLVM). It is a rigorous and validated non-parametric neural model. The FPGA-based computational platform can be configured for either offline or online analysis of the neural ensemble spiking activity, utilizing up to second order GLVM and the neuronal firing data collected from real-word measurement during animal experiment. For the offline neuroinformatics applications, the system can work to efficiently accelerate the estimation process of the GLVM coefficients or to facilitate the model selection procedure (which is an important stage of model inputs reduction), achieving up to thousand-fold speedup compared to the software-based platform. For the online neuroprosthetic applications, the FPGA-based hardware platform is able to perform more consistent cycle-accurate real-time prediction of neural spiking activity. The software platform, although functionally validated in experimental settings, is not capable to guarantee the hard real-time requirement which is crucial for future implantable applications. The differences in calculated results between software and hardware are negligibly small as the normalized mean square error (NMSE) between the two data sets is successfully controlled at the 10-11 scale. After the stage fast model prototyping, we will move from the FPGA-based reconfigurable architectural design to the application specific integrated circuit (ASIC) implementation of the generalized Laguerre-Volterra neural model. In the meantime, advanced design paradigms are going to be adopted and the hardware platform shall continue being upgraded for incorporation of more new features, such as ultra-low power consumption, fault-tolerance and runtime dynamic reconfiguration. Critical issues such as bio-compatibility, packaging, and personalized correct implantation techniques shall be considered for the eventual tap-out, manufacturing and marketing of the prosthetic ASIC.

    Research areas

  • Adaptive computing systems, High performance computing, Neuroinformatics, Neural stimulation, Prosthesis