Optical parallel processing as a basis for neural computation

Jesper Glückstad

Research output: Book/ReportPh.D. thesisResearch

Abstract

The possibility of using opto-electronic hardware as a basis for parallel distributed processing is investigated. The technology is attractive in allowing high density interconnection schemes between two-dimensional arrays of optical sources, modulators and detectors to be achieved. The approach taken in the thesis is to provide a survey of the different ways of implementing optical parallel processing by categorizing the architectures with respect to their interconnection schemes. A natural division is performed between the so-called space variant and space invariant architectures. Architectures that exploit the inherent characteristics of optics such as the concept of space invariance is considered in both theoretical and experimental demonstrations. Novel space invariant processing architectures for nonlinear filtering of gray level images are demonstrated and verified in experimental setups. An extended space invariant scheme capable of performing nonlinear preprocessing of noisy textured images is also considered and demonstrated by the help of software simulations. The parallel preprocessing system is motivated by the way optical information is processed in the early stages of the mammalian visual system. The space invariant processing architectures demonstrate the attractive interplay that can be achieved when interconnections are implemented by the optics and nonlinear point-operations are performed by the electronics. To overcome the problems of misalignment and intensity distortions inherent in the feedback-loop processing arrangements novel and robust self-stabilizing schemes are developed and demonstrated experimentally. Interconnection schemes based on space variant multifacet architectures are considered and analysed in both static, partly adaptive and in fully adaptive versions. Novel ways of implementing partly adaptive and fully adaptive neural training algorithms in compact opto-electronic architectures are demonstrated. A novel proposal for implementing a very simple, robust and dynamic array illuminator based on phase-only transformations is described. An array illuminator provides a way of converting an incident laser beam into an array of bright spots that can be very useful for illuminating arrays of photonic switching devices such as smart pixels, bistable elements and opto-electronic thresholding devices. Alternative illumination schemes are also considered, such as an adaptive "mesh grid" illuminator useful for 3D robot vision applications, and systems based on resonator configurations. The proposed methods are motivated by recently demonstrated self-phase modulation effects in the nonlinear material Bacteriorhodopsin.
Original languageEnglish
Publication statusPublished - Feb 1994
Externally publishedYes

Fingerprint Dive into the research topics of 'Optical parallel processing as a basis for neural computation'. Together they form a unique fingerprint.

Cite this