Introduction to CMMs and AURA based systems

The following pages describe research related to binary neural networks and CMMs undertaken at York. The research is broken into the main themes, which link to details of the projects and research papers related to the projects.

Introduction.

The thrust of CMM based systems is to explore how biologically plausible neural networks can be used for practical computing tasks. AURA is a set of methods based on binary neural networks in the form of correlation matrix memories (CMMs) for high performance pattern matching. The systems are typically aimed at very large datasets (more than 1,000,000 items). The basic CMM structure is a simple biologically plausible network that has the advantage of simple hardware implementation and that is amenable to mathematical analysis.

A CMM has the following basic properties:

• Ability to compress data into a small space, allowing efficient use of computer memory (such as in the address matching research).
• Ability to train on-line, i.e. using one pass through the data. This allows its use in on line applications where little time is available to train a typical neural network.
• Ability to compose the neural networks, in a modular form, into large processing systems. This is undertaken in most applications, but most notably in the CNNAP parallel architecture.
• Simple hardware implementation, as shown by our work on the CORTEX-1 and CORTEX-2 systems. This arises through the use of binary inputs and binary weights.
• Simple to analyse the properties of the network, giving expected performance in terms of storage and speed.

AURA has scored many successes over the last few years, most notably:

• Implementation of a high performance Molecular Matching System in collaboration with GSK and EPSRC.
• Implementation of a postal address matching system in collaboration with The Post Office research group.
• Construction of the Cortex-1 neural network associative supercomputer.
• Construction of a trade mark image matching system.
• Development of a 3D face recognition system with Cybula Ltd.

Overview of research into AURA

In an attempt to organise the research undertaken in CMM based systems we have grouped the research into the main themes as described in the following.

Theoretical analysis of CMMs

This covers the storage properties of CMMs showing how they have good capacity.

Implementation of CMM based systems

This covers both software and hardware implementations of CMMs.

Applications of AURA systems

This contains many applications of AURA methods, some which purely use CMM systems, others that mix CMMs with other weighted neural and statistical methods.