Neural Networks I: Integrated Photonics in Neural Networks
Alex YaSha Yi, Univ. of Michigan, USA
Xingjun Wang, Peking Univ., China
Anthony Choi, Hong Kong Univ., Hong Kong
Integrated Photonics is enabling artificial intelligence (AI). Combination of photonics and AI for photonics-enabled applications is an exciting new prospect. Artificial neural networks (ANNs) constitute the core information processing technology in the fields of artificial intelligence and machine learning, which have witnessed remarkable progress in recent years, and they are expected to be increasingly employed in real-world applications. ANNs are computational models that mimic biological neural networks. They are represented by a network of neuron-like processing units interconnected via synapse-like weighted links. In particular, integrated photonic devices using reservoirs based on physical phenomena has recently attracted increasing interest in many research areas. Various physical systems, substrates, and devices have been proposed for realizing ANNs. A motivation for physical implementation of reservoirs is to realize fast information processing integrated photonic devices with low learning cost. In contrast, physical implementation of reservoirs can be achieved using a variety of physical phenomena in the real world, because a mechanism for adaptive changes for training is not necessary. Actually, integrated photonics is one of the candidates of unconventional computing paradigms based on novel hardware. Although design principles for conventional ANNs, have been examined comprehensively, the following issues require further investigation: how to design physical reservoirs for achieving high computational performance and how much computational power can be attained by individual physical RC systems. Integrated photonics applied to AI has become a more important topic today and we believe the proposed session will attract significant number of attendance. The audience not only comes from photonics community itself, but also from other AI related communities. Conference goals are two fold, one is to let our photonics community to be aware of this emerging direction, the other is to provide a platform for both integrated photonics and AI communities to discuss some future directions which will need experts in both areas to work together.
Neural Networks II: Photonics in Neural Networks: Emerging Concepts
Jeffrey M. Shainline, National Institute of Standards and Technology, USA
As more computing applications require dynamic learning and artificial intelligence, new demands are placed on computational hardware. Opportunities to utilize light are emerging that are quite different from the opportunities present in digital computing and communications. Photonic technologies have much to offer in these emerging application spaces due to the high speed of optical communication and optoelectronic devices, the ability to achieve high signal fan-out, and the massive parallelism possible with optical signals. Ongoing and emerging efforts to use light in adaptive, intelligent systems span many technological platforms, including: photonic neural networks based on delay systems for reservoir computing; deep learning with light propagating through feed-forward neural networks; compound-semiconductor excitable lasers as spiking neurons; silicon photonics for passive routing as well as synaptic weighting with microring weight banks; phase-change materials for optical nonlinearities; and cryogenic spiking neural networks using silicon light sources and single-photon communication. This symposium will cover various roles light may play in emerging hardware for machine learning and neuromorphic computing, new application spaces that may drive demand for photonic or optoelectronic solutions, and the challenges associated with producing commercially viable systems that outperform conventional approaches.
Sonia Buckley, NIST, USA
Progress in Superconducting Optoelectronic Networks for Neuromorphic Computing
Shanhui Fan, Stanford University, USA
Photonic Neural Network: Training, Nonlinearity, and Recurrent Systems
Ken-ichi Kitayama, Osaka University, Japan
Photonic Accelerator : Concept and Perspective
Demetri Psaltis, Ecole Polytechnique Federale de Lausanne, Switzerland
The Past and Future of Optical Neural Networks
Neural Networks III: Recent Advances and Applications of Optical Neural Networks
Volker J. Sorger, George Washington Univ., USA
Thomas Van Vaerenbergh, Hewlett Packard Labs, USA
Nick Usechak, US Air Force Research Laboratory, USA
Photons are ideal information carriers in distributed non von-Neumann compute systems and information processors such as Neural Networks. Artificial neural networks (ANN) are computational network models inspired by signal processing in the brain. These models have dramatically improved performance for many machine-learning tasks, including speech and image recognition. However, today’s computing hardware is inefficient at implementing neural networks, in large part because much of it was designed for von Neumann computing schemes. Significant effort has been made towards developing electronic architectures tuned to implement ANNs that exhibit improved computational speed and accuracy. However, there are a variety of possible implementation options with pro/cons that are being explored in the field today. This symposium reviews and showcases latest approaches to tackle applications using these pioneering directions of optical neural networks. It reviews, highlights, and critically reflects on key technology applications in the fields of data centers, network optimization, image processing, bio-medical diagnostics, or RF information processing or deep-surveillance. The aim of this symposium is to cross-connect hardware engineering with application-driven innovations serving a cross-disciplinary audience. The symposium goals are a) to review and show-case key applications driving the field of optical hardware development, and b) map-out application-based limits of scaling-vectors, and c) connect a multi-disciplinary audience.