- Info
Towards
Ultra-Scalable Neuron Architectures
PhD
RESEARCH PROJECT SYNOPSIS
Artificial
Neural Networks (ANNs) have great potential to
revolutionise modern day technology. The drive
toward systems with thousands of neurons has
highlighted the issue of scalability and power
consumption as a significant limitation as we
move toward the next step, that of millions of
neurons and beyond. This project therefore
seeks to develop novel ideas to support highly
scalable neural architectures in hardware,
which have both a low-power characteristic and
a high interconnect scalability. These
attributes should be beneficial initially for
FPGAs, but also in VLSI implementations, where
transistors and wires scale at different
rates.
The
project will examine two key neural network
approaches. The first will be traditional
fixed-topology network architecture, the
second will be flexible dynamically rewireable
general-purpose architecture. The project will
investigate both continuous valued and spiked
neural networks based upon temporal-serial
spike chains.
Work
will include developing training schemes for
both networks, building interfaces to
high-level machine learning toolkits such as
TensorFlow and Theano, evaluating their
functionality and efficiency, and refining the
hardware models before implementation on FPGA
and in VLSI toolsets. Trade-offs between the
two approaches and their relative suitability
for FPGA vs VLSI will be examined. The core
research question will be to determine if the
temporal spike-chain method can achieve higher
performance (according to various metrics)
than the traditional approach when implemented
in hardware.
The
candidate would ideally have a background
knowledge of VLSI tools, current knowledge of
FPGA toolsets, and an interest in neural
networks and their theory. Students with a
strong Computer Science background and an
interest in digital electronics, or an
electronics student with strong mathematical/
and software skills would be well suited to
the position.
The
project will be supervised by Dr Chris
Crispin-Bailey (Advanced Architectures Group),
and Dr Suresh Manandhar (Artificial
Intelligence Group).