Progress of Optical Neural Networks (Invited)

Bojie Dong; Xiaoyu Li; Yichi Zhang; Bohao Zhang; Zixuan Wang; Wenya Gao; Yanyu Gao; Qi Jia; Xiaoxin Li; Bojian Shi; Yanxia Zhang; Rui Feng; Yongyin Cao; Fangkui Sun; Weiqiang Ding

doi:10.3788/AOS251162

Acta Optica Sinica, Volume. 45, Issue 17, 1720012(2025)

Progress of Optical Neural Networks (Invited)

Bojie Dong¹, Xiaoyu Li¹, Yichi Zhang^1,2, Bohao Zhang^1,3, Zixuan Wang⁴, Wenya Gao¹, Yanyu Gao¹, Qi Jia¹, Xiaoxin Li¹, Bojian Shi¹, Yanxia Zhang¹, Rui Feng¹, Yongyin Cao¹, Fangkui Sun¹, and Weiqiang Ding^1,5、*

Author Affiliations

¹Institute of Advanced Photonics, School of Physics, Harbin Institute of Technology, Harbin 150001, Heilongjiang , China

²Department of Optical Science and Engineering, School of Information Science and Technology, Fudan University, Shanghai 200433, China

³Department of Optica and Optical Engineering, School of Physical Sciences, University of Science and Technology of China, Hefei 230026, Anhui , China

⁴Faculty of Computing, Harbin Institute of Technology, Harbin 150001, Heilongjiang , China

⁵Collaborative Innovation Center of Extreme Optics, Shanxi University, Taiyuan 030006, Shanxi , China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(13)

Fig. 1. Concepts and development of optical neural networks. (a) Development timeline of ONN^{[8-9,11-13,22,35-36]}; (b) classification and basic devices of ONN^{[22,24,37-38]}; (c) basic model of single neuron in neural network, including weighted and nonlinear activation (x₁, x₂, and x_n represent input, w₁₁, w₂₁, and w_n₁ represent weights, f represents nonlinear function, and y₁ represents output)

Download full size

Fig. 2. Multiplexing techniques in ONN. (a) ONN based on WDM^[25,63]; (b) ONN based on OAM-multiplexing^[56] and PDM^[18]; (c) optical RNN based on time-multiplexing^[64-65]

Download full size

$Classification and implementation methods of optical MVM. (a) Principle of MVM for diffractive optical neural network; (b)‒(d) diffractive MVMs based on SLMs[17], metasurface[79] and on-chip diffractive structure[80]; (e) principle of MVM for on-chip integrated optical neural network; (f)‒(h) on-chip MVM based on MZI[81], MRR[82] and PCM[27]$

Fig. 3. Classification and implementation methods of optical MVM. (a) Principle of MVM for diffractive optical neural network; (b)‒(d) diffractive MVMs based on SLMs^[17], metasurface^[79] and on-chip diffractive structure^[80]; (e) principle of MVM for on-chip integrated optical neural network; (f)‒(h) on-chip MVM based on MZI^[81], MRR^[82] and PCM^[27]

Download full size

Fig. 4. Optical nonlinearity implementation in ONN. (a) Nonlinear ONN based on SHG^[84]; (b) nonlinear ONN based on EIT^[85]; (c) ONN based on nonlinear coding system^[89]; (d) ONN based on nonlinear coding of input and system^[67]

Download full size

Fig. 5. In silico training ONN. (a) On-chip integrated ONN^[22]; (b) DONN^[13]; (c) DONN based on programmable metasurface^[98]; (d) tunable OCONN based on semiconductive waveguide^[26]

Download full size

Fig. 6. In situ training ONN based on error BP algorithm. (a) Principle of in situ training by error BP^[99]; (b) PNN based on hybrid training^[103]; (c) DONN based on in situ training by error BP^[100]; (d) OCONN based on based on in situ training by error BP^[101]

Download full size

Fig. 7. Other in situ training methods. (a) ONN with fully forward in situ training based on conjugated relation^[104]; (b) optical spiking neural network with in situ training^[23]; (c) ONN with fully forward in situ training based on positive and negative data^[105]; (d) ONN with fully forward in situ training based on perturbation^[107]

Download full size

Fig. 8. Application of ONN in image processing. (a)(b) Optical convolution for machine vision^[109-110]; (c) optical image scaling^[111]; (d) OCONN for on-chip image processing^[28]

Download full size

Fig. 9. Optical computing acceleration. (a) On-chip photonic tensor core^[66]; (b) Taichi chip^[11]; (c) universal photonic artificial intelligence computing chip^[12]; (d) photonic chip with ultralow latency^[114]

Download full size

Fig. 10. Other applications of ONN. (a) ONN for modes sorting^[117]; (b) ONN for quantum simulation^[37]; (c) ONN for logic operations^[123]; (d) ONN for complex-valued computing^[123]

Download full size

Table 1. Common nonlinear effects and nonlinear coding methods

View table

Table 1. Common nonlinear effects and nonlinear coding methods

Optical nonlinearity	Realization	Threshold / (W/cm²)	Platform	Model	Application
Nonlinear $X_{l}$ effect	Kerr effect^[92]	>1000	Si gating	$X_{l} = p_{3} ({\hat{W}}_{l} \cdot X_{l - 1})$	Dynamic focusing lens
Nonlinear effect	SHG^[84,93]	>1000	LN media	$X_{l} = p_{2} ({\hat{W}}_{l} \cdot X_{l - 1})$	Large-scale computing
Nonlinear effect	Kerr effect^[88]	600^*2	TiN/Al₂O₃ well	$X_{l} = p_{3} ({\hat{W}}_{l} \cdot X_{l - 1})$	Optical differentiation
Nonlinear effect	EIT^[85,94]	0.03^*3	⁸³Rb MOT	$X_{l} = e x p [a A \cdot ({\hat{W}}_{l} \cdot X_{l - 1} {+ b)}^{- 1}]$	Ising phase classification
Nonlinear effect	SA^[87]	0.035^*4	InGaAs well	$X_{l} = p_{n} ({\hat{W}}_{l} \cdot X_{l - 1})$	Iris classification
Nonlinear effect	PL ^[86]	0.000064	Quantum dots	$X_{l} = R e L U ({\hat{W}}_{l} \cdot X_{l - 1} - A)$	Nonlinear convolution
Nonlinear coding	Input coding^[91]	—	SLM	$X_{l} = {\hat{W}}_{l} \cdot e x p (j X_{l - 1})$	Handwriting recognition
Nonlinear coding	System coding^[90]	—	Reflection cavity	$X_{l} = {\hat{W}}_{l} (X_{l - 1}) \cdot X_{l - 1}$	Multimodal recognition
Nonlinear coding	Hybrid coding^[67]	—	VCSEL	$X_{l} = e x p (j {\hat{W}}_{l}) \cdot e x p (j X_{l - 1})$	Handwriting recognition

Table 2. Comparison of image processing performance of partial optical neural networks

View table

Table 2. Comparison of image processing performance of partial optical neural networks

Source	Framework	Classification performance	Reconstruction performance	Application
Shi et al.^[109]	DONN	93.47%^*1	—	Optical image convolution
Zheng et al.^[110]	DONN	98.60%^*2	—	Optical image convolution
Bai et al.^[111]	DONN	—	0.934, 0.979^*4	Optical image zooming
Wu et al.^[28]	OCONN	73.50%^*3	11.69 dB^*5	End-to-end image processing

Table 3. Comparison of performance between optical artificial intelligence (AI) chips and electronic AI chips

View table

Table 3. Comparison of performance between optical artificial intelligence (AI) chips and electronic AI chips

Source	Energy efficiency / (TOPS/W)	Total neurons	Tunable neurons	Scale	Task performacne
Shen et al.^[22]	2.07*	16	16	64	4-category vowel classification
Feldmann et al.^[25]	0.5	64	64	29186	Large-scale matrix multiplication
Xu et al.^[66]	1.27	—	—	—	10-category MNIST classification
Ashtiani et al.^[26]	2.90	67	67	67	4-category letter classification
Xu et al.^[11]	160	4256	160	13.96×10⁶	1000-category classification
Hua et al.^[114]	2.38	64	64	>4095	Ultralow latency matrix multiplication
Ahmed et al.^[12]	1.60	128	128	>16383	Universal AI acceleration
NIVIDIA H100 NVL^[115]	0.17	—	—	—	General purpose computing

Tools

Get Citation

Copy Citation Text

Bojie Dong, Xiaoyu Li, Yichi Zhang, Bohao Zhang, Zixuan Wang, Wenya Gao, Yanyu Gao, Qi Jia, Xiaoxin Li, Bojian Shi, Yanxia Zhang, Rui Feng, Yongyin Cao, Fangkui Sun, Weiqiang Ding. Progress of Optical Neural Networks (Invited)[J]. Acta Optica Sinica, 2025, 45(17): 1720012

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category: Optics in Computing

Received: May. 28, 2025

Accepted: Jul. 10, 2025

Published Online: Sep. 3, 2025

The Author Email: Weiqiang Ding (wqding@hit.edu.cn)

DOI:10.3788/AOS251162

CSTR:32393.14.AOS251162

Topics

laser devices and laser physics

Lasers and Laser Optics

Laser physics

laser manufacturing

Instrumentation, Measurement and Metrology