AISB event Bulletin Item
CALL FOR PARTICIPATION: What Architecture for Neural Hardware?, 26th October, LONDON
Title: What Architecture for Neural Hardware? Speaker: Alex Rast, School of Computer Science, University of Manchester Wednesday 26th October - 16:00-17:30 - Room 343, Huxley Building, South Kensington campus, Imperial College, London, SW7 2A, UK
Abstract: Dedicated hardware is becoming increasingly essential to simulate emerging very large scale neural models. However, the question of what the appropriate architecture is for such neural hardware must take into account the fact that different research groups who might wish to use the hardware may have very different objectives. Previous generations of "neuroprocessor" and "neuromorphic" chips, tended either to hardwire a specific model into the chip, or offer no real advantage over conventional digital processing, meaning as a consequence their interest rarely extended beyond a narrow target audience (usually collaborators of the chip designers). A new approach: the "neuromimetic" architecture, maintains the neural optimisation of dedicated chips while offering FPGA-like universal configurability. As a leading example of this emerging architecture, SpiNNaker is a parallel multiprocessor employing an asynchronous event-driven model with configurable dedicated hardware on the chip to support real-time neural simulation. This makes it capable of supporting multiple models of the neural dynamics, possibly operating simultaneously within the same system. Implementing these models on-chip uses an integrated library-based tool chain that allows a modeller to input a high-level description and use an automated process to generate an on-chip simulation. Results from simulation demonstrate SpiNNaker's ability to support multiple heterogeneous neural models at reasonable scale. SpiNNaker's asynchronous virtual architecture permits greater scope for model exploration, with scalable levels of functional and temporal abstraction, than conventional (or neuromorphic) computing platforms. The neuromimetic architecture opens an intriguing possibility that makes it a compelling choice: using the hardware to establish useful abstractions of biological neural dynamics that could lead to a functional model of neural computation. Biography: Alex Rast is a Research Associate with the SpiNNaker Group at the University of Manchester. He received his Ph.D. from the University of Manchester in 2010 for work on model libraries for configurable neural systems. Prior to joining the University of Manchester he worked at Inficom, Inc., a startup company doing research into advanced processing and communications technologies. His current research interests include extending and standardising neural model libraries including classical models such as the MLP, tool development for neural hardware, parallel and alternative hardware architectures, and programming tools for parallel systems.