Home About Snape Signals Signal
Processing Pattern Recognition Data Mining Research Papers
·
Recognition of User Behaviour
Automatic recognition of the behaviour
an individual has applications ranging from fraud detection to personalisation
of human – machine interaction. Examples of user behaviour are patterns of
financial transactions or a set of TV programmes watched by an individual over
a period of several days. The former may be an indicator of fraud, whilst the
latter could be used to personalise an electronic TV program guide. Snape
Signals has worked as a partner in a DTI funded project called PUMA that has
developed methods of recognising behavioural patterns, and inferring the
identity of the person with which they are associated. Other partners in the
project were BTexact, Imagination Technologies, and the University of
Birmingham. The project examined approaches based on Bayesian and transform
methods, and showed that Latent Semantic Analysis is particularly effective for
behaviour recognition. For more details see: Paper on User
Recognition (166kB)
·
Optimal Pattern Completion
A common task in pattern recognition
is completion of a mapping function given only fragments of the complete
function. Two examples are completion of rule sets, and completion of
continuous non-linear mappings such as used in controllers. Artificial neural
nets are often used for these tasks, but analysis shows that they are generally
only suited to applications in which the mapping is smooth. This is typical of measurements taken from
the physical world which are governed by various forms of inertia. It is not
typical of abstract or symbolic attributes or measurements. An example is the
mapping between attributes such as postcode, and eligibility for a particular
insurance deal. Work by Snape Signals
has focussed on defining the strengths and weaknesses of conventional
artificial neural networks for different applications. This work has been
developed to show how Fourier learning can be applied to completion of both
Boolean and continuous functions. For more detail see: Paper on
Optimal Pattern Completion (434kB)
·
Speaker Recognition
Speaker
recognition systems are often based on the use of MelCeptra extracted from
speech and used within speaker independent and speaker dependent HMMs. The
ratio of the probabilities of the observed speech given by the two models is
used as a measure of likelihood for the particular speaker. Snape Signals has
worked with a specialist speech recognition company to develop linear and
non-linear transform techniques that are applied to the MelCepstra to make the
speaker recognition process more robust. For more details see: Speaker
Recognition Papers (20kB)
Home About Snape Signals Signal
Processing Pattern Recognition Data Mining Research Papers