Transcript Neural Networks Coursework

Combining Multiple Modes of
Information using Unsupervised
Neural Classifiers
Khurshid Ahmad,
Matthew Casey,
Bogdan Vrusias, Panagiotis Saragiotis
http://www.computing.surrey.ac.uk/ncg/
Neural Computing Group,
Department of Computing,
School of Electronics and Physical Sciences,
University of Surrey
1
Content
• Report on preliminary experiments to:
– Attempt to improve classification through
combining modalities of information
– Use a modular co-operative neural network
system combining unsupervised learning
techniques
• Tested using:
– Scene-of-crime images and collateral text
– Number magnitude and articulation
2
Background
• Consider how we may improve classification
through combination:
– Combining like classifiers (e.g. ensemble systems)
– Combining expert classifiers (e.g. modular systems)
• Concentrate on a modular approach to combining
modalities of information
– For example, Kittler et al (1998):
• Personal identity verification using frontal face, face profile
and voice inputs
3
Multi-net Systems
• Concept of combining neural network systems has
been discussed for a number of years
– Both ensemble and modular systems
– Ensemble more prevalent
• Term multi-net systems has been promoted by
Sharkey (1999, 2002) who recently advocated the
use of modular systems
– For example, mixture-of-experts by Jacobs et al 1991
4
Multi-net Systems
• Neural network techniques for classification tend
to subscribe to the supervised learning paradigm
– Ensemble methods
– Mixture-of-experts
• Exceptions include Lawrence et al (1997) and
Ahmad et al (2002)
• Unsupervised techniques give rise to problems of
interpretation
5
Self-organised Combinations
• Our approach is based upon the
combination of different Hebbian-like
learning systems
• Hebb’s neurophysiological postulate (1949)
– Strength of connection is increased when both
sides of the connection are active
6
Self-organised Combinations
• Willshaw & von der Malsburg (1976)
– Used Hebbian learning to associate patterns of activity
in a 2-d pre-synaptic (input) layer and a 2-d postsynaptic (output) layer
– Pre-synaptic neurons become associated with postsynaptic neurons
• Kohonen (1997) extended this in his Selforganising Map (SOM)
– Statistical approximation of the input space
– Topological map showing relatedness of input patterns
– Clusters used to show classes
7
Self-organised Combinations
• Our architecture builds further on this using the
multi-net paradigm
• Can be compared to Hebb’s superordinate
combination of cell assemblies
• Two SOMs linked by Hebbian connections
– One SOM learns to classify a primary modality of
information
– One SOM learns to classify a collateral modality of
information
– Hebbian connections associate patterns of activity in
each SOM
8
Self-organised Combinations
.
.
.
Primary
Vector
.
.
.
Primary
SOM
Bi-directional
Hebbian Network
Collateral
SOM
Collateral
Vector
• SOMs and Hebbian connections trained
synchronously
9
Self-organised Combinations
• Hebbian connections associate
neighbourhoods of activity
– Not just a one-to-one linear association
– Each SOM’s output is formed by a pattern of
activity centred on the winning neuron for the
primary and collateral input
• Training complete when both SOM
classifiers have learned to classify their
respective inputs
10
Classifying Images and Text
Class
Body
Single objects
(close-up)
Primary Image
Collateral Text
Full length shot of body
Nine millimetre browning
high power self-loading pistol
11
Classifying Images and Text
• Classify images based upon images and texts
• Primary modality of information:
– 66 images from the scene-of-crime domain
– 112-d vector based upon colour, edges and texture
• Collateral modality of information:
– 66 texts describing image content
– 50-d binary vector term frequency analysis
• 8 expert defined classes
• 58 vector pairs used for training, 8 for testing
12
Training
•
•
•
•
Image SOM: 15 by 15 neurons
Text SOM: 15 by 15 neurons
Initial random weights
Gaussian neighbourhood function with initial
radius 8 neurons, reducing to 1 neuron
• Exponentially decreasing learning rate, initially
0.9, reducing to 0.1
• Hebbian connection weights normalised
• Trained for 1000 epochs
13
Testing
• Tested with 8 image and text vectors
– Successful classification if test vector’s winner
corresponds with identified cluster for class
• Image SOM:
– Correctly classified 4 images
• Text SOM:
– Correctly classified 5 texts
14
Testing
• For misclassified images
– Text classification was determined
– Translated into image classification via Hebbian
activation
• Similarly for misclassified texts
• Image SOM:
– Further 3 images classified out of 4 (total 7 out of 8)
• Text SOM:
– Further 2 texts classified out of 3 (total 7 out of 8)
15
Comparison
• Contrast with single modality of
classification in image or text SOM
• Compared with a single SOM classifier
– 15 by 15 neurons
– Trained on combined image and text vectors
(162-d vectors)
– 3 out of 8 test vectors correctly classified
16
Classifying Number
• Classify numbers based upon (normalised) image
or articulation?
• Primary modality of information:
– Magnitude representation of the numbers 1 to 22
– 66-d binary vector with 3 bits per magnitude
• Collateral modality of information:
– Articulation representation of the numbers 1 to 22
– 16-d vector representing phonemes
• 22 different numbers to classify
• 16 vector pairs used for training, 6 testing
17
Training
•
•
•
•
Magnitude SOM: 66 by 1 neurons
Articulation SOM: 16 by 16 neurons
Initial random weights
Gaussian neighbourhood function with initial
radius 33 (primary) and 8 (collateral) neurons,
reducing to 1 neuron
• Exponentially decreasing learning rate, initially
0.5
• Hebbian connection weights normalised
• Trained for 1000 epochs
18
Testing
• Tested with 6 magnitude and articulation vectors
– Successful classification if test vector’s winner
corresponds with identified cluster for class
• Magnitude SOM:
– Correctly classified 6 magnitudes
– Magnitudes arranged in a ‘number line’
• Articulation SOM:
– Similar phonetic responses, but essentially
misclassified all 6 articulations
19
Testing
• For misclassified articulation vectors
– Magnitude classification was determined
– Translated into articulation classification via
Hebbian activation
• Articulation SOM:
– 3 articulation vectors classified out of 6
– Remaining 3 demonstrate that Hebbian
association not sufficient to give rise to better
classification
20
Comparison
• Contrast with single modality of classification in
magnitude or articulation SOM
• Compared with a single SOM classifier
– 16 by 16 neurons
– Trained on combined magnitude and articulation
vectors (82-d vectors)
– Misclassified all 6 articulation vectors
– SOM shows test numbers are similar in ‘sound’ to
numbers in the training set
– Combined SOM does not demonstrate ‘number line’
and cannot capitalise upon it
21
Summary
• Preliminary results show that:
– Modular co-operative multi-net system using
unsupervised learning techniques can improve
classification with multiple modalities
– Hebb’s superordinate combination of cell assemblies?
• Future work:
– Evaluate against larger sets of data
– Further understanding of clustering and classification in
SOMs
– Further explore linkage of neighbourhoods, more than
just a one-to-one mapping, and theory underlying
model
22
Acknowledgements
• Supported by the EPSRC Scene of Crime
Information System project (Grant
No.GR/M89041)
– University of Sheffield
– University of Surrey
– Five UK police forces
• Images supplied by the UK Police Training
College at Hendon, with text transcribed by Chris
Handy
23
References
Ahmad, K., Casey, M.C. & Bale, T. (2002). Connectionist Simulation of Quantification Skills.
Connection Science, vol. 14(3), pp. 165-201.
Jacobs, R.A., Jordan, M.I. & Barto, A.G. (1991). Task Decomposition through Competition in a
Modular Connectionist Architecture: The What and Where Vision Tasks. Cognitive Science,
vol. 15, pp. 219-250.
Hebb, D.O. (1949). The Organization of Behavior: A Neuropsychological Theory. New York:
John Wiley & Sons.
Kittler, J., Hatef, M., Duin, R.P.W. & Matas, J. (1998). On Combining Classifiers. IEEE
Transactions on Pattern Analysis and Machine Intelligence, vol. 20(3), pp. 226-239.
Kohonen, T. (1997). Self-Organizing Maps, 2nd Ed. Berlin, Heidelberg, New York: SpringerVerlag.
Lawrence, S., Giles, C.L., Ah Chung Tsoi & Back, A.D. (1997). Face Recognition: A
Convolutional Neural Network Approach. IEEE Transactions on Neural Networks, vol. 8(1),
pp. 98-113.
Sharkey, A.J.C. (1999). Combining Artificial Neural Nets: Ensemble and Modular Multi-Net
Systems. Berlin, Heidelberg, New York: Springer-Verlag.
Sharkey, A.J.C. (2002). Types of Multinet System. In Roli, F. & Kittler, J. (Ed), Proceedings of
the Third International Workshop on Multiple Classifier Systems (MCS 2002), pp. 108-117.
Berlin, Heidelberg, New York: Springer-Verlag.
Willshaw, D.J. & von der Malsburg, C. (1976). How Patterned Neural Connections can be set up
by Self-Organization. Proceedings of the Royal Society, Series B, vol. 194, pp. 431-445.
24
Combining Multiple Modes of
Information using Unsupervised
Neural Classifiers
Khurshid Ahmad,
Matthew Casey,
Bogdan Vrusias, Panagiotis Saragiotis
http://www.computing.surrey.ac.uk/ncg/
Neural Computing Group,
Department of Computing,
School of Electronics and Physical Sciences,
University of Surrey
25
Multi-net Systems
Combination
Decision
Top-down
Static
Multi-net
Systems
Bottom-up
Dynamic
Co-operative
Combination
Mechanism
Competitive
Bottom-up
Combination
Method
Components
Ensemble
Hybrid
Modular
(Fusion)
Sharkey (2002) – Types of Multi-net System
26