Thursday, 09 March,
In addition to the known high-bandwidth benefit, photonics offers two key functionalities with relevance to AI accelerators for machine learning, namely, multiplication and accumulation such as enabled via, for example, modulators and photodetectors, respectively. However, photonics is challenged to provide end-to-end neural network solutions reflected by the challenge of a nonlinear activation function leading to OEO conversions. This channels realisations of deep neural network architectures in the optical domain requiring electronics introducing parasitic conversions. In addition optical accelerators being analog compute engines may require (depending on the application) digital-to-analog domain crossing which is expensive.
In this panel we will review the state-of-the-art in photonic AI accelerators and will project challenges and solutions into the future for photonic and hybrid accelerators for AI and machine intelligence. Here our perspective is open to application spaces in network edge AI and to machine learning training in the cloud.
Glenn Bartolini, Coherent Corp, USA
Nikos Pleros, Aristotle University of Thessaloniki, Greece
Volker Sorger, George Washington University, USA
Xian Xiao, Hewlett Packard Labs, USA
Darius Bunandar, Lightmatter, USA
Hamed Dalir, Optelligence LLC , USA
Johannes Feldmann, Salience Labs, UnitedKingdom
Michael Hochberg, Luminous, USA
Bahram Jalali, UCLA, USA
Francesca Parmigianni, Microsoft Research Cambridge, UnitedKingdom