PIC: Hybrid Photonic-Electronic Reprogrammable Reservoir Computing with Polarization Modes-enhanced Dimensionality

University Of California-San Diego, La Jolla CA

Investigators

Abstract

The recent success of Machine Learning methods based on brain-inspired Neuromorphic Computing (NC) to perform complex information processing tasks spiked significant research in new unconventional computational schemes such as Recurrent Neural Networks (RNNs) and RNN-based Reservoir Computing (RC) which are capable to implement parallel data processing to overcome limitations of conventional sequential computing. Particularly, decomposing the reservoir into an inner network with static weights and an output neurons layer with adaptive and trainable weights allows realization of physical RC where optical-based RC platforms are attractive due to the “speed-of-light” propagation, inherent parallelism, relatively low operation power, and the possibility to harness additional degrees of freedom such as polarization and wavelength. Furthermore, on-chip Photonic Integrated Circuit (PIC) offer enhanced light-matter interaction and modes polarization for enlarged reservoir, and interconnection with CMOS compatible electronics for power efficient electrical reprogrammable feedback. The proposed physical RC PICs are expected to impact mobile applications such as unmanned autonomous vehicles (UAV) and robotic platforms by reducing the need for communication with remote computers, thus avoiding latency and prolonging battery life. (technical description) To realize the reservoir computing (RC) processor utilizing the polarization degrees of freedom on a photonic integrated chip (PIC), we propose the following objectives: (1) numerical and theoretical study aiming to explore the effect of introducing polarization as a new degree of freedom on RC efficiency depending on the underlying architecture of PIC with the electronic feedback elements providing dynamics control; (2) design, fabricate and characterize silicon PIC interconnected with external electronic feedback, admitting the designed architectures; (3) experimentally test the PIC system with external electronic feedback to realize reprogrammable RC tasks, validate the theoretical study and evaluate its performance for relevant applications providing higher accuracy and lower energy consumption compared to state-of-the-art. Rapid prototyping and testing will be performed at UCSD with full scale runs performed at the AIM Photonics foundry. The proposed research is transformative in nature as it will: (i) greatly expand the limits of applicability of RC in CMOS compatible PIC platforms, (ii) develop a fundamental understanding on the effect of reprogrammability on the induced reservoir dynamics and the corresponding performance error, (iii) expand the current notions of both RC and the optical degrees of freedom employed for RC (e.g., polarization). The transformative broader impact of the project arises from the creation of a new much faster and more efficient RC PIC accelerator that will impact mobile applications such as UAV and robotic platforms. The project will provide scientific training for students at graduate and undergraduate levels as well as serve as a basis for outreach, education and collaborative efforts with middle and high schools. Engagement of students from broad backgrounds in Science, Technology, Engineering and Mathematics (STEM) will be continued via the ongoing RET, REU, and COSMOS activities. The program will continue developing a plug & play Integrated Photonics Education Kit (IPEK) and disseminate it to other institutions to implement hands-on classes for a large number of students, and workforce population in the U.S. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.