PhD thesis available

[Stephen Marsland]

Hi,

my PhD thesis
On-line Novelty Detection Through Self-Organisation, with Application to
Inspection Robotics is available at
http://www.cs.man.ac.uk/~marslans/pubs.html

Stephen

Abstract:

Novelty detection, the recognition that a perception differs
in some way from the features that have been seen previously,
is a useful capability for both natural and artificial
organisms. For many animals the ability to detect novel
stimuli is an important survival trait, since the
new perception could be evidence of a predator, while for
learning machines novelty detection can enable useful
behaviours such as focusing attention on novel features, selecting
what to learn and~-- the main focus of this thesis~-- inspection
tasks.

There are many places where an autonomous mobile inspection robot
would be useful~-- examples include sewers, pipelines and even outer
space. The robot could explore its environment and highlight potential
problems for further investigation. The challenge is to have the robot
recognise the evidence of problems. For inspection applications it is
better to err on the side of caution, detecting potential faults that
are, in fact not problems, rather than missing any faults that do
exist. However, by training the robot to recognise each individual
fault, other problems will be missed. This is where novelty detection
is useful. Instead of training the robot to recognise the faults, the
robot learns a model of the `normal' environment that does not have
any problems and the novelty filter detects deviations from this
model.

In training the robot it may well be found that the initial
training set was deficient in some way, for example some
feature that should be found normal was missing and is therefore
always detected as novel. To deal with this situation the novelty
filter should be capable of continuous on-line learning, so that
the filter can learn to recognise the missing feature without having
to relearn every other perception.

This thesis introduces a novelty filter that is suitable for the
inspection task. The novelty filter uses a model of the biological
phenomenon of habituation, a decrement in behavioural response to a
stimulus that is seen repeatedly without ill effects, together with an
unsupervised neural network that learns the model of normality. A
variety of neural networks are investigated for suitability as the
basis of the novelty filter on a number of robot experiments where a
robot equipped with sonar sensors explores a set of corridor
environments.

The particular needs of the novelty filter require a self-organising
network that is capable of continuous learning and that can increase
the number of nodes in the network as new perceptions are seen during
training. A suitable network, termed the `Grow When Required' network,
is devised. The network is applied to a variety of problems, initially
non-novelty detection classification tasks, at which its performance
compares favourably to other algorithms in terms of accuracy and speed
of learning, and then a series of inspection problems~-- both robotic
and not~-- again with promising results. In addition to the sonar
sensors that were used for the earlier robotic inspection tasks, the
output of a CCD camera is also used as input. Finally, an extension to
the novelty detection algorithm is presented that enables the filter
to store multiple models of a variety of environments and to
autonomously select the best one.  This means that the filter can be
used in a set of environments that demonstrate different
characteristics and can automatically select a suitably trained
filter.