Recent Updates:
I have been interested in the immune system metaphor for a while
now.The central idea is to draw inspiration from the attractive
computational features found in the natural immune system.Like many
biologically inspired systems it is adaptive, distributed and
autonomous. Artificial immune systems (AIS) can be applied to
computer security, data mining, optimisation tasks and other domains.
HP are involved in the ARTIST network
of excellence and I am personally involved in the EPSRC funded Danger project
AIS and
Danger
An intriguing idea, the Danger Theory, is causing waves in the
immunological community.It says that the immune system doesn't
care about self and non self, but only about danger and non danger.
HP are industrial partners on an
EPSRC adventure fund programme, a collaboration between
immunologists, computer scientists and security experts, where some
of these implications will be explored. The main thrust of the
project is contained in our ICARIS 2003
paper. I have also co-written a framing
paper which summarises the theory and speculates on its
implications for AIS research. And my AIS overview
for Bristol contains brief notes on the danger theory (among other
things).
Gene
Libraries
Most AIS implementations are based around a canonical algorithm such
as clonotypic learning, which we may think of as individual, lifetime
learning. Yet a species also learns. Gene libraries are often thought
of as a biological mechanism for generating combinatorial diversity
of antibodies. However, they also bias the antibody creation process,
so that they can be viewed as a way of guiding the lifetime learning
mechanisms. Over time, the gene libraries in a species will evolve to
an appropriate bias for the expected environment (based on species
memory). Thus, gene libraries are a form of meta-learning which could
be useful for AIS. This work is an ongoing project with the
University of Bristol and UWE and has resulted in publications at
ICARIS
2005 and ICARIS
2006
AIS and semantic
classification
AIRS is a particularly successful AIS based supervised learning
algorithm. Using a combination of content and semantic features, we
can boost its performance on metadata rich datasets(such as
hierarchical document repositories). Our hypothesis is that this
machine learning approach will be an ideal match for the task of
ontology matching,a task where rigid logic based techniques may fall
foul of real world messiness. The vision is described in this
paper
(accompanying presentation
and poster)
and formed the basis for Julie Greensmith's MSc thesis. Julie
stayed at HP until Christmas 2003 to extend this work.
Julie's work was carried forward by the 2004 BICAS student,
Anastasia Krithara from Bristol University. The focus shifted to
semantic features, and we used other machine learners in addition to
AIRS. The project has been completed, and the findings published in
an HP Labs
Tech Report
AIS and
semantic query
The 'mutate and refine' metaphor used by the AIS is a
promising candidate for query expansion on the semantic web. Perhaps
this approach might provide a novel and efficient way for users to
explore faceted datasets. A Birmingham student, Rana Kashif Ali
investigated this using a dataset drawn from SWED. The project was completed in
Autumn 2004 and a Tech
Report submitted.
Other
Projects
AIS and document classification: Using a
co-evolutionary immune system metaphor,it is possible to implement an
immune based concept learner. This learner has been shown to perform
well both on standard machine learning datasets,and on web page
classification tasks, where the goal is to assess the relevance of a
new web page. The work formed the basis of an MSc
thesis by Jamie Twycross,and a short version is available as an
HP Labs
Tech Report (presented at IIPWM2003).
AIS as a Recommender:My first AIS project used the
immune system metaphor for a collaborative filtering problem. I
implemented an immune system and applied it to the well known
EachMovie recommendation task, achieving pleasing results. The system
is described in a CEC 2002
paper, and analysed in a paper presen
ted to ICARIS2002. A PowerPoint
presentation is also available.
Immune System Resources
I have given a series of presentations about the immune system,the
most recent of which (Feb 2006) is a AIS overview
given to Bristol University. This is an updated version of my
presentation to UCL (Feb
2003). It contains a general overview, details on the document
classification project and brief notes about advanced concepts (eg
idiotypic effects & danger theory). My CEC2002
presentation contains details concerning AIS as a recommender
system.
Please see my resources
page for details on code
availability, useful links
and a list of publications.
|
printable version
