Palmer, Richard Patrick;
(1992)
A novel architecture for a high performance low complexity neural device.
Doctoral thesis (Ph.D), UCL (University College London).
Preview |
Text
A_novel_architecture_for_a_hig.pdf Download (13MB) | Preview |
Abstract
This thesis presents a novel architecture on which to implement neural networks and simple digital signal processing models in a compact and low cost manner. The need for this architecture was identified from a particular application; this application requiring a portable real-time device with which to perform neural network and simple digital signal processing models on speech signals. The development of the architecture was driven by this, and other similar applications, and did not start with any pre-conceived ideas. This allowed complete freedom to be expressed in its development, enabling a more novel architecture to be developed. In this thesis the development of this architecture is reported. This extends from the analysis of current neural models, the identification of the necessary features from these, and the creation of a unified architecture to provide these features. The design of a silicon chip based on this architecture is then presented. The architecture that has been developed combines the programmability and high efficiency of MIMD arrays, with the low complexity in both communications and control of SIMD arrays. The device built using this architecture incorporates a single processing element that consists of only 5300 transistors. Projections indicate that a totally integrated processor array could be built; incorporating fourteen such processors with on-chip memory, A/D and D/A converters. An analysis of the architecture, and other architectures that fit into the same classification class, is then presented. This analysis has been able to identify some common features with MIMD/SIMD machines, and to conclude with some projections for the future.
| Type: | Thesis (Doctoral) |
|---|---|
| Qualification: | Ph.D |
| Title: | A novel architecture for a high performance low complexity neural device |
| Open access status: | An open access version is available from UCL Discovery |
| Language: | English |
| Additional information: | Thesis digitised by ProQuest. |
| Keywords: | Applied sciences; Digital signal processing |
| URI: | https://discovery.ucl.ac.uk/id/eprint/10107623 |
Archive Staff Only
 |
View Item |
