Pruning and optimization of optical neural network as a binary optical trigger
Optical neural networks implemented with Mach-Zehnder Interferometer (MZI) arrays are a promising solution to enable fast and energy-efficient machine learning inference, yet finding a practical application has proven challenging due to sensitivity to thermal noise and loss. To leverage the distinct...
Saved in:
| Main Authors: | , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Frontiers Media S.A.
2025-01-01
|
| Series: | Advanced Optical Technologies |
| Subjects: | |
| Online Access: | https://www.frontiersin.org/articles/10.3389/aot.2024.1501208/full |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Optical neural networks implemented with Mach-Zehnder Interferometer (MZI) arrays are a promising solution to enable fast and energy-efficient machine learning inference, yet finding a practical application has proven challenging due to sensitivity to thermal noise and loss. To leverage the distinct advantages of integrated optical processors while avoiding its shortcomings given the current state of optical computing, we propose the binary optical trigger as a promising field of application. Implementable as small-scale application-specific circuitry on edge devices, the binary trigger runs binary classification tasks and output binary signals to decide if a subsequent energy intensive system should activate. Motivated by the limited task complexity, constrained area and power budgets of binary triggers, we perform 1) systematic, application-specific hardware pruning by physically removing specific MZIs, and 2) application-specific optimizations in the form of false negative reduction and weight quantization, as well as 3) sensitivity studies capturing the effect of imperfections in real optical components. The result is a customized MZI-mesh topology, MiniBokun Mesh, whose structure provides adequate performance and robustness for a targeted task complexity. We demonstrate in simulation that the pruning methodology achieves at least 50% less MZI usage compared to Clements and Reck meshes with the same input size, translating to at least between 4.6% and 24.2% savings in power consumption and a 40% reduction in physical circuitry footprint compared to other proposed unitary MZI topologies, sacrificing only 1%–2% drop in inference accuracy. |
|---|---|
| ISSN: | 2192-8584 |