Wave-based Integrated Nonlinear Generative System for Artificial Intelligence

Projektbeschreibung

Photonikhardware für flexible Definitionen von KI-Funktionalitäten

Das rasante KI-Wachstum treibt die Nachfrage nach leistungsfähigerer und energieeffizienterer Rechenhardware voran. Mit Unterstützung durch das Marie-Skłodowska-Curie-Maßnahmenprogramm wird Projekt WINGS-AI eine neue Photonikhardware-Klasse entwickeln, die die Parallelität neuronaler Netze mithilfe kompakter, energieeffizienter nichtlinearer Photonikchips und innovativer modellfreier Trainingsmethoden nutzt. Im Rahmen des Projekts werden programmierbare Bausteine in einem nichtlinearen optischen Substrat auf Lithiumniobat-Basis geschaffen. Diese Fortschritte werden die direkte Implementierung von KI-Funktionen in der Hardware ermöglichen, was zu neuronalen Netzen auf der physikalischen Ebene führt, deren Operationen über ein kontinuierliches nichtlineares optisches System verteilt sind.

Ziel

Artificial intelligence (AI) is disrupting nearly every aspect of modern life, including powerful generative models like ChatGPT which are capable of generating human-like text and images. However, the prevalence of AI comes at a growing cost that is clearly unsustainable, and new solutions for AI hardware are urgently needed. The aim of my project is to directly address this challenge by developing a new class of photonic hardware – specifically a Wave-based Integrated Nonlinear Generative System for Artificial Intelligence (WINGS-AI) to implement the inherent parallelism of neural networks in a low energy, low footprint integrated nonlinear photonic chip, paired with innovative and model-free training approaches that account for fabrication non-idealities.
The key physical concept underlying the project is to implement fundamental building blocks in a non-structured, spatially continuous yet fully programmable nonlinear optical substrate based on a Lithium Niobate photonic chip. This is a major conceptual shift when compared to the current state of the art for integrated photonic AI which is focused on (and limited by) the use of pre-fabricated lithography-based topologies. I will be able to near-arbitrarily define the functionality and transformations of AI computational elements as training and optimization targets of the continuous medium, yielding a physical layer NN with operations distributed along a continuous, nonlinear dynamical system.
I will perform this work at the CNRS institute FEMTO-ST (France) with primary host Dr. Daniel Brunner and at Cornell University (USA) with secondment host Prof. Peter McMahon whose laboratory has recently demonstrated the programmable nonlinear chip in a non-AI context. Combining the expertise of both institutions creates a unique opportunity to advance next-generation photonic generative AI. This project WINGS-AI would profoundly aid me reaching professional maturity as an independent PI in this cutting-edge field.