University of the Aegean
Photonic neuromorphic processors based on semiconductor lasers' dynamics for reservoir computing and spiking neural networks
Abstract
dc:descriptionThis doctoral thesis explores the development of photonic neuromorphic processors which aim to address three critical challenges posed by Big Data and Internet of Things: the unprecedented needed processing rates, the requirement for low power consumption and minimal hardware. At first, bio-inspired primitives such as Artificial neural networks were chosen to tackle these challenges. However, the simultaneous pursuit of two fundamental but conflicting objectives compromised their efficiency as it was imperative to increase the number of neurons-nodes and synapses-connections and to reduce the network complexity and power consumption simultaneously. In this context, neuromorphic computing schemes aimed to mitigate power consumption but still struggle to address the connectivity and scalability challenges of large-scale networks. Time Delay Networks (TDNs) offer a solution to this impediment by multiplexing neurons and synapses in the temporal domain, thus minimizing hardware requirements. However, they suffer from significant speed penalties, which limit their application in real-time tasks. In order to lift these restrictions, this thesis exploits the inherent merits of high-speed photonics so as to maximize the processing capabilities of TDNs. This attempt starts with the successful implementation of a photonic TDN based on quantum dot spin Vertical-Cavity Surface-Emitting Laser (QD sVCSEL). The proposed system demonstrated a fourfold increase in neuron count, whereas more importantly matched the accuracy of state-of-the-art neuromorphic schemes, utilizing a fraction of neurons. Additionally, the use of multiple masks and the injection of the non modulated field significantly improved accuracy, allowing for less sophisticated hardware designs. Then, a reconfigurable photonic TDN processor using Fabry-Perot (FP) lasers was investigated both numerically and experimentally. Towards this direction, the first step consisted of investigating experimentally the ability to increase virtual neurons count and performance by utilizing multiple longitudinal modes of the FP. After that, the capability of the proposed FP-TDN to further accelerate processing without sacrificing accuracy was numerically demonstrated via a new Spectro-temporal multiplexing (SPTM) technique. The FP-TDN processor assisted with SPTM was able to achieve a Bit Error Rate of 10-4 in a transmission impairment mitigation task, with feedback of only 140ps whereas HD-FEC compatibility was demonstrated for feedbacks of even 60ps. In addition, the FP-TDN recorded 95.95% accuracy on the MNIST task simultaneously unlocking processing rates of 255.1Mimages/sec. The next part includes a time-multiplexing neuromorphic scheme, but in this case entailing spiking nodes. In particular, a deep photonic spiking convolutional neural network was introduced using VCSEL nodes and exploiting time multiplexing so as to regulate neuron counts from 62 to 2020, depending on processing rate demands. It demonstrated unsupervised learning capabilities with high noise immunity and achieved processing rates which ranged from 720 ns to 5 ns per image. Finally, a spiking VCSEL based TDN was developed. The processor was capable of addressing various classification tasks, such as Iris, MNIST and an Imaging Flow Cytometry task whereas it demonstrated ultra low power consumption at the processing stage (3.23 pJ per image). For the first time, this work unified neuromorphic computing and neuromorphic sensing under a common photonic platform, establishing a pathway for future research into scalable, energy-efficient photonic processors for real-time applications.
Degree
thesis:*- Grantor dc:publisher
- University of the Aegean
- Year dc:date
- 2024
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Skontranis, Menelaos
- Σκοντράνης, Μενέλαος
Subjects
dc:subject × 16- Φωτονική
- Υπολογιστική χρονικής καθυστέρησης
- Ακραία μάθηση χρονικής καθυστέρησης
- Spiking νευρωνικά δίκτυα
- Νευρομορφική υπολογιστική
- Photonics
- Time delay reservoir computing
- Time delay extreme learning machine
- Spiking neural networks
- Neuromorphic computing
- Άλλες Επιστήμες Μηχανικού και Τεχνολογίες
- Επιστήμες Μηχανικού και Τεχνολογία
- Μηχανική και Τεχνολογίες, άλλοι τομείς
- Other Engineering and Technologies
- Engineering and Technology
- Engineering and Technologies, miscellaneous
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/58262