Advanced Processor Technologies Home
APT Advanced Processor Technologies Research Group

Scalability and Robustness of Artificial Neural Networks

Evangelos Stromatias

Abstract

Artificial Neural Networks (ANNs) appear increasingly and routinely to gain popularity today, as they are being used in several diverse research fields and many different contexts, which may range from biological simulations and experiments on artificial neuronal models to machine learning models intended for industrial and engineering applications. One example is the recent success of Deep Learning architectures (e.g., Deep Belief Networks [DBN]), which appear in the spotlight of machine learning research, as they are capable of delivering state-of-the-art results in many domains.While the performance of such ANN architectures is greatly affected by their scale, their capacity for scalability both for training and during execution is limited by the increased power consumption and communication overheads, implicitly posing a limiting factor on their real-time performance. The on-going work on the design and construction of spike-based neuromorphic platforms offers an alternative for running large-scale neural networks, such as DBNs, with significantly lower power consumption and lower latencies, but has to overcome the hardware limitations and model specialisations imposed by these type of circuits. SpiNNaker is a novel massively parallel fully programmable and scalable architecture designed to enable real-time spiking neural network (SNN) simulations. These properties render SpiNNaker quite an attractive neuromorphic exploration platform for running large-scale ANNs, however, it is necessary to investigate thoroughly both its power requirements as well as its communication latencies.This research focusses on around two main aspects. First, it aims at characterising the power requirements and communication latencies of the SpiNNaker platform while running large-scale SNN simulations. The results of this investigation lead to the derivation of a power estimation model for the SpiNNaker system, a reduction of the overall power requirements and the characterisation of the intra- and inter-chip spike latencies. Then it focuses on a full characterisation of spiking DBNs, by developing a set of case studies in order to determine the impact of (a) the hardware bit precision; (b) the input noise; (c) weight variation; and (d) combinations of these on the classification performance of spiking DBNs for the problem of handwritten digit recognition. The results demonstrate that spiking DBNs can be realised on limited precision hardware platforms without drastic performance loss, and thus offer an excellent compromise between accuracy and low-power, low-latency execution. These studies intend to provide important guidelines for informing current and future efforts around developing custom large-scale digital and mixed-signal spiking neural network platforms.

The thesis is available as PDF (27 MB).