University of the Aegean
Study of photonic neuromorphic devices based on spiking quantum dot lasers and reconfigurable photonic synapses for signal processing applications
Abstract
dc:descriptionThis doctoral thesis focuses on the investigation of photonic devices for the design a high-speed and power efficient neuromorphic photonic platform for spiking neural networks (SNN). Each SNN consists of two major components, namely the spiking neuron and the synapses that provide the interconnection. Transferring this topology in the photonic domain, semiconductor quantum dot lasers (QD) are investigated as spiking nodes, since they are described by dynamics that are similar to those of the biological neural cells, but they produce spikes in the form of optical pulses and at time-scales that are 6 orders of magnitude faster. Synapses are implemented with coherent photonic meshes of tunable Mach Zehnder interferometers, providing matrix-vector multiplication with passive elements at speeds limited only by the available electronics (GHz rates).QD lasers are theoretically studied for the case of neuromorphic nodes based on optical injection. The theoretical description considers the dual-state competent dynamics that are unique to the QD lasers. By using elements from bifurcation theory, the neurocomputational properties of QD lasers are uncovered for both lasing states. The numerical evaluations reveal the QD laser under this topology can operate both as an integrate and fire (IF) and a resonate and fire (RF) neuron, thus offering high-speed rate encoding and frequency selectivity. The interplay between the carriers in the ground state (GS) and the first excited state (ES), provides the ability to generate ES spiking from GS spiking, a property that is crucial due to recent works that link GS lasing to excitations and ES lasing to inhibition. Spike events as fast as 130 picoseconds are calculated, revealing the suitability of this platform for high-speed processing.Apart from the optically injected topology, the two-section topology is also investigated with a gain and a saturable absorber (SA) section. This system is studied experimentally, by deploying both electrical and optical triggering at the photonic neuron. Contrary to conventional two-section lasers, this device offers an additional biasing point for excitability that is associated with the interplay between GS and ES. Bifurcation analysis reveals that different bias conditions lead to both IF and RF functionalities, whereas the co-existence of ES lasing and GS spiking reduces significantly the temporal width of spike events down to 500 picoseconds even for a long 2 mm device.Finally, from the viewpoint of synapses, a Bayesian description of the coherent mesh is presented, that treats each synapse as random Gaussian variable. The training algorithm optimizes the probability density function (PDF) of each synapse. This scheme is numerically evaluated at a non-ideal large-scale mesh that uses thermo-optic phase shifters for image classification. The Bayesian neural network provides information about the robustness of all photonic synapses, which can be harnessed to achieve more than 70 % power reduction at the electronic components that control the synapses, 30 % reduction of the required actuators and robustness to the thermal crosstalk effect. Therefore, this method is capable to offer resilient, low complexity and low power operation at next-generation high-speed synaptic meshes.
Degree
thesis:*- Grantor dc:publisher
- University of the Aegean
- Year dc:date
- 2024
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Sarantoglou, Georgios
- Σαράντογλου, Γεώργιος
Subjects
dc:subject × 22- Κβαντικές τελείες
- Τεχνολογία, Φωτονική
- Τεχνητή νοημοσύνη
- Νευρομορφική υπολογιστική
- Quantum dots
- Photonics technology
- Artificial intelligence
- Neuromorphic computing
- Φυσική
- Φυσικές Επιστήμες
- Οπτική
- Επιστήμη Ηλεκτρολόγου Μηχανικού, Ηλεκτρονικού Μηχανικού, Μηχανικού Η/Υ
- Επιστήμες Μηχανικού και Τεχνολογία
- Υπολογιστές, Υλικό (hardware) και Αρχιτεκτονική
- Μηχανική και συστήματα επικοινωνιών, Τηλεπικοινωνίες
- Physical Sciences
- Natural Sciences
- Optics
- Electrical Engineering, Electronic Engineering, Information Engineering
- Engineering and Technology
- Computer science, Hardware and Architecture
- Communication engineering and systems, Telecommunications
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/57089