Project

Back to overview

Novel Architectures for Photonic Reservoir Computing

English title Novel Architectures for Photonic Reservoir Computing
Applicant Ortega Juan-Pablo
Number 175801
Funding scheme Project funding (Div. I-III)
Research institution Fachbereich für Mathematik und Statistik Universität St. Gallen
Institution of higher education University of St.Gallen - SG
Main discipline Information Technology
Start/End 01.05.2018 - 30.04.2021
Approved amount 600'000.00
Show all

All Disciplines (5)

Discipline
Information Technology
Mathematics
Microelectronics. Optoelectronics
Electrical Engineering
Material Sciences

Keywords (10)

reservoir computing; optical neural networks; volatility forecasting; liquid state machines; integrated optics; brain-inspired computing; neural computing; time series forecasting; echo state networks; Hi-Res EEG classification

Lay Summary (French)

Lead
Ce projet adresse les défis poses par le Big Data par le biais du développement de dispositifs de haute performance à l’aide d’un paradigme d’apprentissage machine récent connu sous l’appellation « Reservoir computing » (RC). Le RC est bien adapté aux applications qui demandent des calculs en ligne et il est un candidat sérieux pour surmonter les limitations des machines de Turing traditionnelles dont les algorithmes tournent sur des architectures de type Von Neumann. En plus, le RC est l’un des rares approches computationnelles qui peuvent être traduits sur du hardware dédié.
Lay summary

L’objectif principal de ce projet est la conception et la fabrication de reservoirs photoniques compacts. La photonique intégrée s’est avérée une approche prometteuse pour la réalisation physique de puces pour le reservoir computing (RC). Dans NAPRECO, nous suivrons deux approches pour la conception de dispositifs compacts qui devraient emmener le hardware pour RC à un nouveau niveau de performance. D’abord, nous étudierons expérimentalement les réseaux de semi-conducteurs amplificateurs optiques (SOAs) pour aller au-delà ce que jusqu’à l’instant a été découvert simplement dans le cadre de simulations. Ensuite, un nouveau concept de RC photonique sera développé à l’aide de la dynamique des patterns interférométriques optiques.

Le développement du hardware sera optimisé par des études théoriques qui fourniront un guidage mathématique fort. D’abord, le concept RC basé sur les interférences demande une modélisation mathématique précise afin d’optimiser son architecture. Ensuite, plusieurs résultats théoriques montrent l’universalité du RC en tant que paradigme computationnel pour des entrées analogiques et digitales, mais les contraintes technologiques peuvent entamer la validité de ces résultats, qui doivent être réévalués dans ces contextes particuliers.

Des applications apprentissage machine spécifiques dans le contexte financier et médical seront explorées. Nous nous focaliserons dans la prédiction de séries temporelles multivariées avec des inputs multifréquence, prédiction de la volatilité financière multivariée et la classification de signaux électroencéphalographiques de haute résolution.


Direct link to Lay Summary Last update: 04.04.2018

Responsible applicant and co-applicants

Employees

Project partner

Abstract

This proposal addresses the challenges of the Big Data era by developing high performance devices in the context of a recently established machine learning paradigm known as reservoir computing (RC). RC is well adapted to applications that require real-time online computing and is a serious candidate to outperform conventional Turing machines whose algorithms run on von Neumann architectures. Additionally, It is one of the few computing approaches that can be directly mapped to dedicated hardware. We articulate the proposed research around three main axes:As dedicated hardware, we will design and fabricate compact photonic reservoirs. Integrated photonics has proved to be a promising approach to physically implement RC chips. In NAPRECO, we will follow two approaches to compact designs that bring hardware RC to a new level of performance. Scaled networks of coupled semiconductor optical amplifiers (SOAs) will be first studied experimentally, going beyond what has so far only been shown in simulations. A novel concept of compact photonic RC will be developed next, based on the dynamics in an optical interference pattern. Inspired by a RC demonstration using surface water waves our photonic equivalent would yield drastically smaller length scales and time constants.In synergy with the development of the hardware concepts, the RC architecture will be optimized from a theoretical perspective. A strong mathematical guidance will be first provided to physically implement the two approaches highlighted. First, the RC concept based on interference requires an accurate mathematical modeling to optimize its architecture. Second, various theoretical results show the universality of RC as a computational paradigm both for digital and analog inputs but mapping RC to a dedicated hardware implies important restrictions that might limit the validity of those results. Finally, we will characterize the class of problems to which our concepts can be applied, quantify their performances, identify possible limitations, and suggest routes for improvement.Specific machine learning applications will be explored in the financial and medical domains. We will focus on multivariate time series forecasting based on mixed frequency explanatory input signals, multivariate financial volatility forecasting, and classification of high-resolution electroencephalographic signals (Hi-Res EEG). Having in common a large volume of data and an unfavorable signal to noise ratio, such applications require a powerful signal processing tool, and will ideally demonstrate the potential of hardware RC solutions to impact relevant economic sectors.NAPRECO is an intrinsically multidisciplinary collaboration that combines state-of-the-art knowledge in the mathematical and experimental realization of RC systems. It will advance the scientific understanding in a complex field, establish Swiss researchers in a world-leading role, and develop a vision towards critical applications. The expertise in integrated photonics and material science is covered by IBM Research-Zurich. The Sankt Gallen group will focus on the theoretical optimization of the reservoir architecture and on the development of the applications, having a well-grounded track record in both tasks. Both applicants have working experience, and are well embedded in international partnerships. Scientific results are therefore expected at a sustained rate and with a significant impact in strategic fields such as macroeconometrics, financial econometrics, and clinical neuroscience.
-