Use this URL to cite or link to this record in EThOS:
Title: Neural pipelines : for the co-ordination of activity in a multi-layered neural network
Author: Naylor, Rebecca Frances
ISNI:       0000 0004 2737 6075
Awarding Body: University of York
Current Institution: University of York
Date of Award: 2011
Availability of Full Text:
Access from EThOS:
Access from Institution:
The `Neural Pipeline' is introduced as an artificial neural network architecture that controls information flow using its own connection structure. The architecture is multi-layered with `external' connections between the layers to control the data. Excitatory connections transfer data from each layer to the next and inhibitory feedback connections run from each layer to the previous layer. Using these connections a layer can temporarily silence the previous layer and stop further inputs until it finishes processing. When excitation and inhibition are balanced, waves of activity propagate sequentially through the layers after each input; this is `correct' behaviour. When the system is `over' inhibited, the inhibitory feedback outweighs the excitation from the input. At least one layer remains inhibited for too long so further inputs cannot stimulate the layer. Over inhibition can be corrected by increasing the delay between inputs. When the system is `under' inhibited the excitation in the layer is larger than the inhibition. The layer is therefore not silenced and continues to spike. In the layers, excitatory and inhibitory spiking neurons are randomly inter-connected. Changing layer parameters influences the system behaviour. Recommendations for correct behaviour include: low neuron connectivity and balancing the external inhibition and layer activity. With variations of only the internal topology and weights, all three behaviours can be exhibited. Each layer is trained as a separate Liquid State Machine, with readout neurons trained to respond to a particular input. A set of six shapes can be learnt by all layers of a three layer Neural Pipeline. The layers are trained to recognise different features; layer 1 recognising the position while layer 2 identifies the shape. The system can cope when the same noisy signal is applied to all inputs, but begins to make mistakes when different noise is applied to each input neuron. The thesis introduces and develops the Neural Pipeline architecture to provide a platform for further work.
Supervisor: O'Keefe, Simon ; Austin, Jim Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID:  DOI: Not available