Multi-dimensional optical neural network
Fuente:
arXiv
Saved in:
| Main Authors: | , , , , , |
|---|---|
| Format: | Preprint |
| Published: |
2024
|
| Subjects: | |
| Online Access: | |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866916494426767360 |
|---|---|
| author | Jia, Zhetao Rubio, Hector Neim, Lilian Park, Jagang Preble, Stefan Kanté, Boubacar |
| author_facet | Jia, Zhetao Rubio, Hector Neim, Lilian Park, Jagang Preble, Stefan Kanté, Boubacar |
| contents | The development of deep neural networks is witnessing fast growth in network size, which requires novel hardware computing platforms with large bandwidth and low energy consumption. Optical computing has been a potential candidate for next-generation computing systems. Specifically, wavelength-division multiplexing (WDM) has been widely adopted in optical neural network architecture to increase the computation bandwidth. Although existing WDM neural networks architectures have shown promise, they face challenges in the integration of light sources and further increase of the computing bandwidth. To overcome these issues, we introduce a mode-division multiplexing (MDM) strategy, offering a new degree of freedom in optical computing within the micro-ring resonator platform. We propose a MDM approach for small-scale networks and a multi-dimensional architecture for large-scale applications, supplementing WDM with MDM to enhance channel capacity for computations. In this work, we design and experimentally demonstrate key components for the MDM computing system, i.e., a multimode beam splitter, a thermo-optical tuner for the high-order mode, and a multimode waveguide bend. We further show a 2-by-2 matrix multiplexing system fabricated in a foundry that works for both MDM and MDM-WDM computing, which confirms that our approach successfully increases the input vector size for computing and ensures compatibility with existing WDM networks. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2411_16140 |
| institution | arXiv |
| publishDate | 2024 |
| record_format | arxiv |
| spellingShingle | Multi-dimensional optical neural network Jia, Zhetao Rubio, Hector Neim, Lilian Park, Jagang Preble, Stefan Kanté, Boubacar Optics The development of deep neural networks is witnessing fast growth in network size, which requires novel hardware computing platforms with large bandwidth and low energy consumption. Optical computing has been a potential candidate for next-generation computing systems. Specifically, wavelength-division multiplexing (WDM) has been widely adopted in optical neural network architecture to increase the computation bandwidth. Although existing WDM neural networks architectures have shown promise, they face challenges in the integration of light sources and further increase of the computing bandwidth. To overcome these issues, we introduce a mode-division multiplexing (MDM) strategy, offering a new degree of freedom in optical computing within the micro-ring resonator platform. We propose a MDM approach for small-scale networks and a multi-dimensional architecture for large-scale applications, supplementing WDM with MDM to enhance channel capacity for computations. In this work, we design and experimentally demonstrate key components for the MDM computing system, i.e., a multimode beam splitter, a thermo-optical tuner for the high-order mode, and a multimode waveguide bend. We further show a 2-by-2 matrix multiplexing system fabricated in a foundry that works for both MDM and MDM-WDM computing, which confirms that our approach successfully increases the input vector size for computing and ensures compatibility with existing WDM networks. |
| title | Multi-dimensional optical neural network |
| topic | Optics |
| url | https://arxiv.org/abs/2411.16140 |