Neuromorphic photonic systems with lasers

Student thesis: Doctoral Thesis

Abstract

In this thesis we investigate the technology of Vertical Cavity Surface Emitting Lasers (VCSELs) as potential candidates for neuromorphic (brain-like) photonic computing hardware elements, towards the future realisation of ultrafast, energy-efficient and light-enabled information processing platforms. VCSELs are a type of semiconductor laser sources that not only exhibit numerous beneficial characteristics, such as low power requirements, compactness, low manufacturing costs and high modulation speeds, but also offer exciting prospects as photonic emulators of biological neurons. Hence, in this thesis we develop and study the behaviour of VCSEL-based artificial photonic neurons and reveal their capability to activate sub-nanosecond neuron-like optical spiking responses (3 orders of magnitude faster than state of the art electronic implementations of biological neurons). We further explore the control of the non-linear neuron-like dynamics in VCSELs before taking advantage of their ultrafast optical spiking signals to produce examples of spike-based photonic information processing systems.First, we study the neuron-like excitable (spiking) dynamics exhibited by VCSELs under external optical injection. Both the activation and inhibition of 100 ps-long (GHz rates) spiking responses is demonstrated using both modulated optical injection and modulated bias current. These mechanisms are shown to elicit neuron-like optical spiking regimes, both controllably and consistently, in VCSELs. The similarities of the responses achieved in the analysed VCSEL neurons to neuronal models is then investigated, where we reveal further the underlying neuron-like behaviours (such as threshold-and-fire, and spiking refractory periods) in these photonic devices.Second, we investigate the networking capability of the developed VCSEL neurons. By building experimental configurations of coupled VCSELs we demonstrate their ability to communicate optical spiking signals. Both the activation and inhibition of optical spikes is shown to be propagated in (1-to-1) feedforward architectures, revealing the output of an artificial VCSEL neuron is cascadable across layers in a network. Further we demonstrate the activation of two downstream VCSEL neurons in a diverging (1-to-2) architecture and create a three layer (1-to-1-to-1) VCSEL network inspired by biological cell layers in the retina. The latter is achieved without signal manipulation between network layers, with the implementation of all optical signals using commercially-sourced VCSELs.Finally, we discuss the successful application of VCSEL neurons in functional neuromorphic photonic information processing demonstrations. We achieved digital-to-spike conversion of return-to-zero (RZ) and non return-to-zero (NRZ) signals for the interfacing of spiking neuromorphic platforms with traditional digital technologies. We also explored the time-division multiplexing (TDM) of VCSEL inputs for the creation of a virtual converging (many-to-one) network architecture. Using this technique we revealed for the first time (experimentally) the neuronal integrate-and-fire behaviour of the VCSEL neuron. Exploiting this key neuronal behaviour, we implemented a single artificial VCSEL neuron as processing element, and demonstrated the coincidence detection of fast (sub-nanosecond) optical inputs with the firing of an optical spiking response. Further employing the integrate-and-fire capability, we demonstrated both 4-bit binary pattern recognition and image processing (edge-feature detection) tasks with a single VCSEL neuron. Moreover, we utilised the spike-based edge-feature detection of the VCSEL neuron (alongside a software implemented spiking neural network, SNN) to successfully classify digits from the MNIST handwritten digit database, achieving a high classification accuracy of 96.1%. Successful operation of different information processing tasks was therefore achieved with systems based on VCSEL neurons, that utilised both ultrafast (GHz rate) optical spiking representations and hardware-friendly (commercially-sourced) photonic components. We therefore believe VCSELs, with their exciting characteristic and highlighted neuronal behaviours, serve as excellent potential candidates for neuromorphic photonic implementations of novel ultrafast and efficient information processing systems for brain-inspired computing and light-enabled Artificial Intelligence (AI) hardware.
Date of Award26 Apr 2023
Original languageEnglish
Awarding Institution
  • University Of Strathclyde
SponsorsEPSRC (Engineering and Physical Sciences Research Council)
SupervisorAntonio Hurtado (Supervisor) & Michael Strain (Supervisor)

Cite this

'