Craven, Michael P.
Inter-chip communications in an analogue neural network utilising frequency division multiplexing.
PhD thesis, University of Nottingham.
As advances have been made in semiconductor processing technology, the number of transistors on a chip has increased out of step with the number of input/output pins, which has introduced a communications ’bottle-neck’ in the design of computer architectures. This is a major issue in the hardware design of parallel structures implemented in either digital or analogue VLSI, and is particularly relevant to the design of neural networks which need to be highly interconnected.
This work reviews hardware implementations of neural networks, with an emphasis on analogue implementations, and proposes a new method for overcoming connectivity constraints, by the use of Frequency Division Multiplexing (FDM) for the inter-chip communications. In this FDM scheme, multiple analogue signals are transmitted between chips on a single wire by modulating them at different frequencies.
The main theoretical work examines the number of signals which can be packed into an FDM channel, depending on the quality factors of the filters used for the demultiplexing, and a fractional overlap parameter which was defined to take into account the inevitable overlapping of filter frequency responses. It is seen that by increasing the amount of permissible overlap, it is possible to communicate a larger number of signals in a given bandwidth.
Alternatively, the quality factors of the filters can be reduced, which is advantageous for hardware implementation. Therefore, it was found necessary to determine the amount of overlap which might be permissible in a neural network implementation utilising FDM communications.
A software simulator is described, which was designed to test the effects of overlap on Multilayer Perceptron neural networks. Results are presented for networks trained with the backpropagation algorithm, and with the alternative weight perturbation algorithm. These were carried out using both floating point and quantised weights to examine the combined effects of overlap and weight quantisation. It is shown using examples of classification problems, that the neural network learning is indeed highly tolerent to overlap, such that the effect on performance (i.e. on convergence or generalisation) is negligible for fractional overlaps of up to 30%, and some tolerence is achieved for higher overlaps, before failure eventually occurs. The results of the simulations are followed up by a closer examination of the mechanism of network failure.
The last section of the thesis investigates the VLSI implementation of the FDM scheme, and proposes the use of the operational transconductance amplifier (OTA) as a building block for implementation of the FDM circuitry in analogue VLSI.
A full custom VLSI design of an OTA is presented, which was designed and fabricated through Eurochip, using HSPICE/Mentor Graphics CAD tools and the Mietec 2.4µ CMOS process. A VLSI architecture for inter-chip FDM is also proposed, using adaptive tuning of the OTA-C filters and oscillators.This forms the basis for a program of further work towards the VLSI realisation of inter-chip FDM, which is outlined in the conclusions chapter.
Thesis (University of Nottingham only)
||neural network, learning algorithm, multiplexing, backpropagation, weight perturbation, simulation, filter, oscillator, VLSI, ANN, FDM, text-to-speech
||T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK7800 Electronics
||UK Campuses > Faculty of Engineering > Department of Electrical and Electronic Engineering
||13 Feb 2013 11:57
||19 Sep 2016 21:24
Actions (Archive Staff Only)