DPhil Thesis: A Neurosymbolic Approach to the Classification of Scarce and Complex Data

[Margarita Sordo]

Dear Connectionists,

My DPhil thesis on knowledge-based neural networks for classification
of scarce and complex medical data is now available. 

Title: "A Neurosymbolic Approach to the Classification of
        Scarce and Complex Data" 

It can be found at:

http://www.acl.icnet.uk/lab/aclsanchez.html


Abstract:

Artificial neural networks possess characteristics that make them a
useful tool for pattern recognition and classification. Important
features such as generalization, tolerance to noise and graceful
degradation make them a robust learning paradigm. However, their
performance strongly relies on large amounts of data for training.
Therefore, their applicability is precluded in domains where data are
scarce.

Knowledge-based artificial neural networks (KBANNs)  provide a means
for combining symbolic and connectionist approaches into a hybrid
methodology capable of dealing with small datasets. The suitability of
such networks has been evaluated in binary-valued domain theories.

After replicating some initial results with a binary-valued domain
theory, this thesis presents new results with scarce and complex
real-valued medical data. 31P magnetic resonance spectroscopy (MRS) of
normal and cancerous breast tissues provide good testbeds to assess the
advantages of such a methodology over other, more traditional
connectionist approach for classification purposes in constrained
domains.

Experimental work confirmed the suitability of the proposed
neurosymbolic approach for real-life applications with such
constraints. Knowledge in the symbolic module helps to overcome the
difficulties found by the connectionist module when confronted with
small datasets. Details of breast tissue metabolism and MRS are
presented.  Knowledge acquisition methodologies for gathering the
required knowledge for the definition of the domain theories are also
described. Future directions for improving the KBANN methodology are
discussed.



===============================================================================
Margarita Sordo	Sanchez         		

ms at acl.icnet.uk

Advanced Computation Laboratory
Imperial Cancer Research Fund
61 Lincoln's Inn Fields,
London WC2A 3PX, 
England, United Kingdom

phone: 44 (020) 7242 0200 Ext 2911 
       44 (020) 7269 2911 (direct)
fax:   44 (020) 7269 3186  
===============================================================================