Transcript Can We Count on Neural Networks?

You wanted to know what the Matrix is?
Matthew Casey
[email protected]
The Matrix (1999), © 1999 Warner Bros.
2
Modelling the Action of the Brain
• You wanted to know what the Matrix is?
– ‘What you can feel […] smell […] taste and see […] are simply
electrical signals interpreted by your brain’ (The Matrix 1999)
– The outlook may be dystopian, but in artificial intelligence, the
dream is still to build ‘intelligent machines’ (even if they do take
over the world)
• Even Alan Turing was similarly motivated, for example,
in 1947, when he wrote:
– "I am more interested in the possibility of producing models of
the action of the brain than in the practical applications to
computing.“ (Hodges 1992:363)
• Yet modelling “the action of the brain”
– Requires knowledge of the actions to be modelled
– Requires a robust understanding of the tools used
– Has practical applications as well
Hodges, A. (1992). Alan Turing: The Enigma. London: Vintage, Random House.
3
Knowledge of the Actions
• Why?
– To understand ‘the brain’ better
– To use knowledge of the brain to build more
‘intelligent machines’
• Do we know enough about the brain?
– We have a developing understanding (e.g. Carter
2000)
– …and detailed models of specific aspects (e.g.
Feigenson et al 2004)
– …but our knowledge appears quite poor (e.g.
Olshausen 2005)
Carter, R. (2000). Mapping the Mind. London, UK: Phoenix.
Feigenson, L., Dehaene, S. & Spelke, E. (2004). Core systems of number. Trends in Cognitive Sciences, vol. 8(7), pp. 307314.
Olshausen, B.A. & Field, D.J. (2005). How Close are we to Understanding V1? Neural Computation, vol. 17, pp. 1665-1699.
4
Knowledge of the Tools
• Why?
– To explore new architectures (theory and application)
– To understand how biological systems can give rise to behaviour
• Which architectures and algorithms?
– What types of neuron, network or algorithm?
– Our focus is on combining neural networks (Sharkey 1999)
– Limited theory, but wide application, not just computational
neuroscience (e.g. Kittler at al 1998)
• But we need a sufficiently robust understanding of these
– To understand model behaviour
– To generalise and apply elsewhere
– So far, this robust understanding exists only for a small number of multinet architectures (negative correlation learning and mixture-of-experts)
Sharkey, A.J.C. (1999). Multi-Net Systems. In Sharkey, A. J. C. (Ed), Combining Artificial Neural Nets: Ensemble and
Modular Multi-Net Systems, pp. 1-30. London: Springer-Verlag.
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.
5
Practical Applications
• The grand aim of artificial intelligence?
– To build ‘intelligent machines’?
• Grand Challenges (Hoare & Milner 2004)
– Architecture of Brain and Mind
– ‘Bottom-up specification […] of computational models’
– ‘Top-down development of a new kind of theory’
• But
– Knowledge spread across different disciplines
– Foresight Cognitive Systems (Sharpe 2003): we need
to develop an inter-disciplinary understanding
Hoare, T & Milner, R. (2004). Grand Challenges in Computing: Research. UK Computing Research Committee (UKCRC).
Sharpe, B. (2003). Foresight Cognitive Systems Project: Applications and Impact.
Industry, Office of Science and Technology.
London: Department of Trade and
6
Solutions?
• Modelling the ‘action of the brain’
– Use available tools to prototype aspects of cognition: neural
networks and signal processing, …
– Use inter-disciplinary knowledge: psychophysics and
neurobiology
• Grand Challenge suggests:
– Build increasingly more complex models
– Combine theory and empirical results to build better models
• Most of the brain is dedicated to some form of sensory
processing
– A good place to start…
– …building upon the wealth of computational work that has
already been done, but hasn’t quite solved the problem
7
The Matrix (1999), © 1999 Warner Bros.
8
Multi-sensory Processing
• Could you understand what was being said in the film clip?
– You should be able to, even without the sound
– Your other senses, memory, emotions, etc. work together
• But, can we get a machine to do the same thing?
– Computer vision, speech recognition, etc. are hard tasks
– The brain does it very well – but how?
• Uni-modal or multi-modal processing?
– Typically, only single modalities have been modelled
– Yet evidence suggests that there is some low-level cross-sensory
processing (Thesen et al 2004)
– Can computational models benefit from a similar approach?
• Aim
– To explore multi-sensory processing to see if it can help us build
models/machines that are closer to being intelligent
– We therefore need to build ever more complex models of the brain that
can process different sensory inputs in an integrated way
Thesen, T., Vibell, J.F., Calvert, G.A. & Österbauer, R.A. (2004). Neuroimaging of Multisensory Processing in Vision,
Audition, Touch, and Olfaction. Cognitive Processing, vol. 5(2), pp. 84-93.
9
Multi-modal Processing
• Multi-sensory integration leads to a multi-modal
understanding: the whole ‘picture’
• Numerical cognition
– Exploring a multi-modal understanding in numeracy
– Integrated cognitive abilities for manipulating numbers:
subitization, counting, addition and number representation
(Casey et al, Casey 2004, Ahmad et al 2002)
• Multi-sensory processing and tools
– Combines visual and linguistic inputs to produce a single output
– Combining supervised and unsupervised learning in parallel and
in sequence
– But better (and less specific) models of vision, audition, etc.
needed
Casey, M.C. & Ahmad, K. (accepted). A Competitive Neural Model of Small Number Detection. Neural Networks.
Casey, M.C. (2004). Integrated Learning in Multi-net Systems. Unpublished doctoral thesis. Guildford, UK: University of
Surrey.
Ahmad, K., Casey, M.C. & Bale, T. (2002). Connectionist Simulation of Quantification Skills. Connection Science, vol.
14(3), pp. 165-201.
10
Low-level Vision
• Modelling low-level human vision (with Sowden)
– Category learning task (Notman et al 2005)
– Task dependence tunes low-level processing
– Categorical perception effect: measurable difference in ‘within
class’ versus ‘between class’ discrimination
– What causes this CP Effect?
• Well-established area of investigation and computational
modelling
– However, new understanding of sensory processing: low-level
vision is not static
– Dynamic changes to how we process low-level visual input
depending upon task (visual and task inputs)
– Linked to low-level cross-modal processing
Notman, L.A., Sowden, P.T. & Özgen, E. (2005). The Nature of Learned Categorical Perception Effects: A Psychophysical
Approach. Cognition, vol. 95(2), pp. B1-B14.
11
Categorical Perception
Do these belong to the same or a different category?
12
Category Learning
135o
Category A
180o
Images combine
an f and 3f
grating
225o
90o
Distance changes
through learning
270o
45o
3f phase angle
Notman et al 2005
0o
Category B
315o
13
Receptive Field Modelling
• Modelling low-level vision:
– 2-D Gabor filtering: frequency, phase and orientation (cf. Itti & Koch
2001)
– Split into receptive fields
– Neuron per field, fed into discrimination model
– Task driven: discrimination/categorisation
– Meant to learn how to combine receptive field values
• But…
– Grappling with ‘plausible’ models of vision
– MLP only: needs to model ‘templates’ and lateral inhibition (competitive
learning?)
– Assumes a model of vision that may be wrong (cf. Olshausen 2005)
– Relies upon simplistic grating patterns (as used in human tests)
• Despite the problems
– This simple model is starting to show that the CP Effect can be
reproduced because of the process of learning categories
Itti, L. & Koch, C. (2001). Computational Modelling of Visual Attention. Nature Reviews Neuroscience, vol. 2(3), pp. 194-203.
Olshausen, B.A. & Field, D.J. (2005). How Close are we to Understanding V1? Neural Computation, vol. 17, pp. 1665-1699.
14
What Next?
• Work so far has:
– Built multi-net adaptive models of specific cognitive
abilities
– Focussed on single vision, linguistic and task inputs
• Need to:
– Build better model of vision and other senses
– Need to combine these models
– Perhaps demonstrate via simple robotics
• But…
– We still lack a robust understanding of the tools
– Despite exploring novel architectures and algorithms:
sequential and parallel systems (Casey et al 2004)
Casey, M.C. & Ahmad, K. (2004). In-situ Learning in Multi-net Systems. In Yang, Z.R., Everson, R. & Yin, H. (Ed),
Proceedings of the 5th International Conference on Intelligent Data Engineering and Automated Learning (IDEAL 2004),
Lecture Notes in Computer Science 3177, pp. 752-757. Heidelberg: Springer-Verlag.
15
Work in Progress
• Limited theory on combined systems:
– Ensemble (NCL) and modular systems (mixture-of-experts)
• Individual networks are well understood
– Are multi-nets just single networks?
– Are neural network ensembles just partially connected feedforward
systems (cf. Brown 2004)?
• We need a better understanding of useful architectures (without
getting lost in the detail)
– Ensembles: game theory approach (Zanibbi, Casey & Brown)
– Ensembles: application to classification (Zhang & Casey)
– Recurrence and single-nets (Taskaya-Temizel & Casey)
• …of generic architectures:
– Set theoretic (Shields & Casey)
– Infer properties of combined system from components
• …and to think about other aspects of the brain:
– Emotion (Pavlou & Casey)
Brown, G. (2004). Diversity in Neural Network Ensembles. Unpublished doctoral thesis. Birmingham, UK: University of
Birmingham.
16
Coming Soon…
Workshop on Biologically Inspired
Information Fusion
UniS 22nd and 23rd August 2006
Matthew Casey, Paul Sowden, Tony Browne, Hujun Yin
Bringing together computer scientists, engineers, psychologists and
biologists to discuss multi-sensory processing
http://www.soc.surrey.ac.uk/ias/workshops/biif/
17
Thank you
Questions?
[email protected]
Numeracy and the Brain
Right Hemisphere
Left Hemisphere
Parietal Lobe
Parietal Lobe
Comparison
Frontal
Lobe
Magnitude
Representation
Magnitude
Representation
Comparison
Frontal
Lobe
Verbal
System
Arithmetic
Facts
Visual Number
Form
Temporal Lobe
Occipital Lobe
Visual Number
Form
Temporal Lobe
Dehaene, S. (2000). The Cognitive Neuroscience of Numeracy: Exploring the Cerebral Substrate, the Development, and
the Pathologies of Number Sense. In Fitzpatrick, S.M. & Bruer, J.T. (Eds), Carving Our Destiny: Scientific Research
faces a New Millennium, pp. 41-76. Washington: Joseph Henry Press.
19