Efficient Neuromorphic Photonics Processor Architectures forMachine Learning
Open AccessIn the past decade, photonics and neuromorphic computing have been pioneered as possible alternatives to our current computing systems. Due to its low propagation loss, photonics has become not only a suitable medium for long-distance communications, but also for short intra-chip communications. There has been a significant advance in the photonic integration industry, and photonic devices are finding niche applications in computing. Meanwhile, neuromorphic computers, inspired by the biological brain, have evolved to overcome the problems inherent in von-Neumann-based systems.\parThis research is motivated by the advancements in these fields, which provide high energy efficiency (attojoule/MAC), high speed, and high bandwidth, in addition to potentially more parallelism. A further limitation of the prior research has been that it has concentrated on the device-level and small circuit-level neuromorphic nanophotonics. We exploit this synergy by combining novel concepts from each of the fields. Our primary approach has focused on design and analysis of efficient neuromorphic photonic architectures for machine learning. We researched the design of efficient specialized neuromorphic photonic computing systems for machine learning that,-Can execute state-of-the-art neural networks-Exhibit great energy efficiency and speed -Are programmableWe studied various classes of neural networks. Based on preliminary analysis and experiments, it appeared that convolutional neural networks are well suited to be realized using photonics. Consequently, we concentrated on convolutional neural networks and optimized our design for the convolution operation. In order to achieve further optimization, we investigated the possibility of using mathematical transforms to perform the convolution operation. By using the Winograd transform, we were able to reduce the costs without sacrificing performance. We explored the design space by-Highlighting potentials, pitfalls, and limitations-Developing a design and simulation framework for neuromorphic photonicsToward the end of this study, we also investigated the potential of using photonics for large-scale neuroscience simulations, such as that of the brain. We proposed and analyzed optical communication links as an efficient method of facilitating synaptic connections for brain simulations. We achieved this by,-Modeling a photonic-plasmonic link within a brain simulation tool commonly-used by computational neuroscientists-Performing a comparison study between the optical implementation and its electrical counterpart -Designing and simulating an adaptive network-on-chip augmented with a photonic-plasmonic express interconnect to further improve the speed and energy efficiencyThe results of our study indicate that neuromorphic photonic systems could significantly improve the energy efficiency and speed of current systems. Nevertheless, converting digital and analog signals and interfacing with electrical components consume a substantial portion of those gains. It is likely that future improvements in technology and architecture will reduce these costs.
- All rights reserved
Notice to Authors
If you are the author of this work and you have any questions about the information on this page, please use the Contact form to get in touch with us.
| Thumbnail | Title | Date Uploaded | Visibility | Actions |
|---|---|---|---|---|
|
|
Mehrabian_gwu_0075A_15778.pdf | 2022-03-06 | Open Access |
|