FLOPS: EFficient On-Chip Learning for OPtical Neural Networks Through Stochastic Zeroth-Order Optimization

Gu, Jiaqi; Zhao, Zheng; Feng, Chenghao; Li, Wuxi; Chen, Ray T.; Pan, David Z.

doi:10.1109/dac18072.2020.9218593

Cited by 16 publications

(23 citation statements)

References 33 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…A detailed introduction to ONNs can be found in Appendix A. Beyond offline training [21], ONN on-chip training methods are proposed to offload the process back onto photonics [24,20,17], shown in Table 1. Brute-force device tuning (BFT) [41,58] and evolutionary algorithms [56] are applied to search MZI settings.…”

Section: Related Workmentioning

confidence: 99%

“…An adjoint variable method (AVM) [24] is proposed to directly evaluate gradients using in-situ light field monitoring. Stochastic zeroth-order optimization (ZOO) [20,17] is later applied to improve the training efficiency. However, prior methods are hard to scale to larger ONNs either due to algorithmic inefficiency or unrealistic hardware complexity.…”

Section: Related Workmentioning

confidence: 99%

“…BFT [41] PSO [56] AVM [24] FLOPS [20] [43], forward and feedback matrices [1,38], and input feature maps [36]. Quantization to the pre-activation gradients is adopted in [51] to induce sparsity by trading off quantization steps and performance.…”

Section: Related Workmentioning

confidence: 99%

“…Due to the above two limitations, it is challenging to obtain true derivatives w.r.t. the MZI rotation phases in U and V * [41,20,17], casting fundamental in-situ optimization difficulty as ONN scales up.…”

Section: Related Workmentioning

confidence: 99%

“…However, training on photonic neural chips is non-trivial and much less explored than on conventional platforms. Prior work [56,24,20,17] only demonstrated small prototypes, and their scalability and efficiency are rather limited. To push the limits of DNNs in optics, we propose an efficient three-stage learning framework L 2 ight that consists of variation-agnostic identity calibration, alternate projection-based parallel mapping, and multi-level sparse subspace learning.…”

Section: Introductionmentioning

confidence: 99%

See 4 more Smart Citations

L2ight: Enabling On-Chip Learning for Optical Neural Networks via Efficient in-situ Subspace Optimization

Gu¹,

Zhu²,

Feng³

et al. 2021

Preprint

Self Cite

View full text Add to dashboard Cite

Silicon-photonics-based optical neural network (ONN) is a promising hardware platform that could represent a paradigm shift in efficient AI with its CMOScompatibility, flexibility, ultra-low execution latency, and high energy efficiency. In-situ training on the online programmable photonic chips is appealing but still encounters challenging issues in on-chip implementability, scalability, and efficiency. In this work, we propose a closed-loop ONN on-chip learning framework L 2 ight to enable scalable ONN mapping and efficient in-situ learning. L 2 ight adopts a three-stage learning flow that first calibrates the complicated photonic circuit states under challenging physical constraints, then performs photonic core mapping via combined analytical solving and zeroth-order optimization. A subspace learning procedure with multi-level sparsity is integrated into L 2 ight to enable in-situ gradient evaluation and fast adaptation, unleashing the power of optics for real on-chip intelligence. Extensive experiments demonstrate our proposed L 2 ight outperforms prior ONN training protocols with 3-order-of-magnitude higher scalability and over 30× better efficiency, when benchmarked on various models and learning tasks. This synergistic framework is the first scalable on-chip learning solution that pushes this emerging field from intractable to scalable and further to efficient for next-generation self-learnable photonic neural chips. From a co-design perspective, L 2 ight also provides essential insights for hardware-restricted unitary subspace optimization and efficient sparse training. We open-source our framework at link.Preprint. Under review.

show abstract

Section: Related Workmentioning

confidence: 99%

Section: Related Workmentioning

confidence: 99%

Section: Related Workmentioning

confidence: 99%

Section: Related Workmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 3 more Smart Citations

L2ight: Enabling On-Chip Learning for Optical Neural Networks via Efficient in-situ Subspace Optimization

Gu¹,

Zhu²,

Feng³

et al. 2021

Preprint

Self Cite

View full text Add to dashboard Cite

show abstract

Silicon Photonics for Future Computing Systems

Shafiee

Pasricha

Nikdast

2022

Wiley Encyclopedia of Electrical and Electronics Engineering

View full text Add to dashboard Cite

The primary goal of this article is to provide an overview of silicon photonics technology and its applications in the design and improvement of current and future computing systems. We start by reviewing silicon photonics technology and introducing some of its benefits and challenges as well as providing some background on it. Next, we introduce some fundamental silicon photonic components in the design of silicon photonic integrated circuits (PICs) and optical interconnect for computing systems as well as their operating principles and applications. These components can be active, such as photodetectors and optical modulators, or passive, such as silicon‐on‐insulator (SOI) waveguides. Subsequently, we discuss the application of silicon photonics to improve the communication and computation infrastructure in future computing systems, while reviewing the state‐of‐the‐art and some design and implementation challenges. Finally, we discuss several research opportunities to push forward the application of silicon photonics in the design of future computing systems.

show abstract

Dual Optical Frequency Comb Neuron: Co‐Developing Hardware and Algorithm

Zhang

Tao

Yan

et al. 2023

Advanced Intelligent Systems

View full text Add to dashboard Cite

Previous studies on photonic neural network have demonstrated that algorithm can inspire hardware design. This study seeks to demonstrate that hardware can also inspire algorithm design. To further exploit the advantages of photonic analog computing, the authors develop hardware and algorithm simultaneously for photonic convolutional neural networks. Specifically, this work developed an architecture called dual optical frequency comb neuron (DOFCN) enabled by an integrated microcomb to perform cosinusoidal nonlinear activation and vector convolution without temporal or spatial dispersion and large‐scale modulator arrays. Furthermore, DOFCN‐based composite vector convolutional neural networks (CVCNNs), an optical‐electric hybrid model, are proposed to perform classification and regression tests in signal modulation format identification and optical structure inverse design tasks, respectively. The ablation experiments show that under 4‐bit precision limit, the element‐wise activation CVCNN has 14% higher classification accuracy, 76% lower regression residuals, and 100% higher training efficiency than that of the 32‐bit standard convolutional neural network (CNN). DOFCN exhibits impressive spectral information processing ability to facilitate signal‐processing tasks related to optics and electromagnetics.

show abstract

FLOPS: EFficient On-Chip Learning for OPtical Neural Networks Through Stochastic Zeroth-Order Optimization

Cited by 16 publications

References 33 publications

L2ight: Enabling On-Chip Learning for Optical Neural Networks via Efficient in-situ Subspace Optimization

L2ight: Enabling On-Chip Learning for Optical Neural Networks via Efficient in-situ Subspace Optimization

Silicon Photonics for Future Computing Systems

Dual Optical Frequency Comb Neuron: Co‐Developing Hardware and Algorithm

Contact Info

Product

Resources

About