| EPSRC Reference: |
GR/J38987/01 |
| Title: |
DYNAMIC BEHAVIOUR OF NETWORKS OF SPIKING NEURONS |
| Principal Investigator: |
Taylor, Professor J |
| Other Investigators: |
|
| Researcher Co-Investigators: |
|
| Project Partners: |
|
| Department: |
Mathematics |
| Organisation: |
Kings College London |
| Scheme: |
Standard Research (Pre-FEC) |
| Starts: |
01 September 1993 |
Ends: |
31 August 1995 |
Value (£): |
94,256
|
| EPSRC Research Topic Classifications: |
|
| EPSRC Industrial Sector Classifications: |
|
| Related Grants: |
|
| Panel History: |
|
|
Summary on Grant Application Form |
|
To investigate the dynamic behaviour of models of neural networks of spiking neurons by simulation and mathematical and information-theoretic analysis. The main aim is to provide a mathematical description of simple spiking neural networks which should constitute a solid basis for more detailed future models. The models to be investigated will incorporate the basic elements of the cortical circuitry such as feed-forward, lateral and feedback projections and spiking neurons. The latter are regarded as necessary to overcome some of the inadequacies of the formal neuron models used in other artificial neural networks.Progress:The project work is under four headings:1. Analysis of pRAM networks.2. Leaky Integrate and Fire(LIF) networks.3. Learning in Nets of Spiking neurons.4. Analysis of Nets of Spiking neurons.There has been progress in all of these areas. In the first, and with contribution to the fourth, there has been extension of the feed-forward nets of pRAMs to the case of inhibitory feedback, and the manner in which this type of connectivity, with feedback inhibition, might represent activity in visual cortex underlying the transmission of neural activity from one module to a succeeding one. The manner in which the latency of the transmission of spikes to arrive coincidentally on later neurons depends on the jitter at the first (retinal) layer and on the necessary number of coinciding spikes has been determined both by simulation and by mathematical analysis. The manner in which such processing might explain masking phenomena and the generation of oscillations due to inhibitory feedback has been elucidated.Under the second heading, there has been the development of a multi-module structure which has coupled modules with neurons in one module being LlFs with a much longer time constant than in the other layer. These different modules are being modelled to simulate the semantic and working memory type of modules which might be interpreted as lying in layer 4 and layers 2/3 and 4/5 in cortex. They have feed-forward and feedback connections as in the first topic. They are also structured so as to have receptive fields somewhat mimicking that of the various areas of visual cortex. The development of suitable simultaneous feature detectors in the modules, arising by simple Hebbian learning rules (as under topic 3), with increasing complexity developing the higher in the hierarchy the neurons lie, is presently under analysis, but is expected from mathematical analysis (again under project 4) to allow for such structures to arise. This is being related to the elaborate cells of Tanaka and to the geons of Biederman.Besides the development of the above learning system there has also been considerable progress in creating a pRAM system with the ability to support back-error propagation learning. This has required the use of the non-linearisation of the pRAM in order to give similar structures to the system as to the usual artificial neural network (ANN). At the same time it has been found necessary to extend the pRAM to have addresses for both excitatory and inhibitory inputs, so leading to the pRAM+/-. This approach has allowed a net of pRAMs to be developed to learn by using similar formulae to those for the standard ANN, but now in a directly hardware-implementable form.Further developments have also occurred in the area of classification by pRAMs. A new adaptive classification system has been developed in which output prototypes for input classes are modified by reinforcement learning, so as to optimise the classification accuracy. This is presently under mathematical analysis, as well as being applied to real-world problems in speech analysis and target.
|
|
Key Findings |
|
This information can now be found on Gateway to Research (GtR) http://gtr.rcuk.ac.uk
|
|
Potential use in non-academic contexts |
|
This information can now be found on Gateway to Research (GtR) http://gtr.rcuk.ac.uk
|
|
Impacts |
|
| Description |
This information can now be found on Gateway to Research (GtR) http://gtr.rcuk.ac.uk |
| Summary |
|
| Date Materialised |
|
|
|
|
Sectors submitted by the Researcher |
|
This information can now be found on Gateway to Research (GtR) http://gtr.rcuk.ac.uk
|
| Project URL: |
|
| Further Information: |
|
| Organisation Website: |
|