All-optical associative learning element

Thesis

This thesis examines the idea that Pavlovian associative learning may be a building block in neural networks. With emphasis on hardware acceleration, on-chip optical associative learning device is conceived as an alternative to artificial neural networks (ANNs) hardware for high-speed applications. It is well known that ‘conventional’ ANNs, particularly in the form of modern deep neural networks (DNNs), are usually carried out using the backpropagation method. In the method, network weights are updated such that the net output better resembles the desired output after each forward-backward iteration, until convergence is met. A distinct approach is presented using associative learning as the basis of AI learning process. In formulating the architecture, associative learning is linked with supervised learning, based on their common goal of associating certain inputs with ‘correct’ outputs. The intuition leads to a reductionist framework to the problem, facilitating the transition from a single (or monadic) Pavlovian single Input-Teacher association to any arbitrary n Input-Teacher associations.

The hardware associative learning network is realised using silicon-based on-chip optical waveguides. Optical implementation of the learning network enables simultaneous parallel calculations using wavelength division multiplexing (WDM), which increases the capacity of AI information processing. This is adopted in the work described in this thesis. Monolithic integration of waveguide components, through which optical signals are channeled, augurs well for its adoption as artificial intelligence (AI) hardware accelerator. Machine learning using associative learning hardware network is demonstrated, with the network solving pattern and image recognition tasks. The patterns are randomised optical signals binary pattern, while the images are the discretised 72x72 patterns in the form of image pixels. Computational density of 118 TOPS/mm² is demonstrated, limited only by the available setup in the laboratory. The density is estimated to reach approximately 2.5× 104 TOPS/mm², with 50 GHz modulation using 16 wavelengths. In solving the image recognition task, the hardware network is shown to be able to create a predictive model capable of generalising a set of patterns, instead of recognising only one particular pattern.

The optical associative learning device paves the way for further realisations of photonic intelligent systems that can infer both symbolically and numerically via associations. Further improvements in other relevant metrics (for example, volume and learning energy) are anticipated on different material platform and/or with other optimisation methods. These will be reserved for a future time when they shall have been more completely worked out. The device realisation potentially opens up new avenues of research in machine learning architectures and algorithms.

Authors: Tan, JYS

• DOI: 10.5287/ora-av4oqrg2b
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: Integrated optics; optical computing
• Subjects: Integrated optics, optical computing