Tuesday, 22 May 2012, 7:00pm - 8:30pm

Nigel Sedgwick, Cambridge Algorithmica Ltd.

Pattern Matching plays an increasingly important part of modern life, in which machines are taking on additional roles. Examples include automatic speech recognition, biometric identification of people, monitoring of machines for faults or for automatic timing of maintenance; even radio and line communications.

This talk will describe a generalised approach, with benefits of simplicity and, often, improved performance. The approach is useful for quite large numbers of features, with nonlinear correlations between them; it also suits applications with medium numbers of features, that require more accurate matching than currently.

The approach has more intrinsic tolerance than many, of missing feature measurements, for both training data and for operational classification. It handles this in a unified way with noise modelling. It can incorporate many existing processing techniques, including those that have already been found successful in particular applications.

A key issue for such Pattern Matching is the advent of more powerful computer processors with significantly larger memory. This make practical, techniques in Statistical Pattern Matching that would not previously have been thought attractive.

As well as use with supervised training, the approach can also incorporate manual guidance of modelling to various extents, thereby covering supervised, unsupervised and semisupervised training. The approach is based on discriminative training; it uses improved modelling of Probability Density Functions (PDFs), including going well beyond Gaussian distributions, and even beyond Gaussian mixtures. It is useful for multisensor feature fusion and also for fusion of the outputs of prior multi-algorithmic analysis.

Evaluation of the technique will be shown, particularly using graphical performance measures based on Receiver Operating Characteristic (ROC) curves.

The talk will touch on some interesting though modest parallels with possible mechanisms in human brains, that go beyond those of Artificial Neural Networks.

Download: Slide deck

Download: mp3 Podcast


Steria Limited, Hemel Hempstead
Three Cherry Trees Lane, Hemel Hempstead, HP2 7AH


Free for members and non-members

Last updated 16th November, 2020 at 4:11pm