Progressive Training Scheme for Recognition Error of Optical Neural Networks

Fig. 1. Structural composition diagram of MZI-ONN. (a) MZI-ONN structure; (b) rectangularly arranged MZI array-based optical interference unit; (c) typical 2×2 MZI

Download full size

Fig. 2. Average fidelity of RM of different scales when beam splitter error and phase shifter error vary independently. (a) 4×4 RM; (b) 6×6 RM; (c) 8×8 RM; (d)16×16 RM

Download full size

Fig. 3. Recognition accuracy of 6×6×3 MZI-ONN structure with component error for MNIST dataset

Download full size

Fig. 4. Recognition accuracy of MZI-ONN at different scales under different beam splitter errors

Download full size

Fig. 5. Simulation results of 4×4×3 MZI-ONN. (a) Accuracy and loss in Iris dataset after normal training and PT scheme optimization; (b) confusion matrix with component error; (c) confusion matrix after PT scheme optimization

Download full size

Fig. 6. Simulation results of 6×6×3 MZI-ONN. (a) Accuracy and loss in MNIST dataset after normal training and PT scheme optimization; (b) confusion matrix with component error; (c) confusion matrix after PT scheme optimization

Download full size

Fig. 7. Recognition accuracy at different MIZ-ONN scales

Download full size

Table 1. Progressive training algorithm

View table

Table 1. Progressive training algorithm

Algorithm 1： progressive training algorithm
Input： input features N， maximum iterations M， phase before normal training for each iteration $D = \{d_{1}, d_{2}, \dots, d_{M}\}$ ， phase after normal training for each iteration P $= \{p_{1}, p_{2}, \dots, p_{M}\}$ Output： best phase $p_{M}$
1 for m=1 to M
2 d= $d_{m}$ ； p= $p_{m}$ ； k=int（m+1/100）
3 if k%2==0 and k≠0 then
4 c=（2N-2）*k/2； d=c-N+2
5 else if k%2≠0 and k≠0 then
6 c=（2N-2）*（k+1）/2； d=c-N
7 else
8 c=0； d=0
9 end if
10 p［0：c］=d［0：c］； w=range（d， c）
11 if （m+1）%100==0 and m≠0 then
12 for i in w do
13 P［i］= round （1）
14 end for
15 else
16 break from line 2
17 end if
18 end for
19 return $p_{M}$

Table 2. Comparison of different optimization schemes

View table

Table 2. Comparison of different optimization schemes

Training scheme	Networksize	Dataset	Number of MZI	Number of detectorused to detect error	Accuracy improvement /（percentage point）
GA training^［13］	4	Iris	6	1	23.10
GA training^［13］	8	MNIST	28	1	32.40
Progressive training	4	Iris	6	1	64.15
	8	MNIST	28	1	37.00
	16	MNIST	128	1	36.25
RRM^［10］	16	MNIST	256	1	Matrix error decreases by five orders of magnitude
Hardware correction^［12］	16	MNIST	128	128	Matrix error decreases by an order of magnitude

Tools

Get Citation

Copy Citation Text

Pengxing Guo, Zhengrong You, Weigang Hou, Lei Guo. Progressive Training Scheme for Recognition Error of Optical Neural Networks[J]. Acta Optica Sinica, 2024, 44(7): 0720001

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category: Optics in Computing

Received: Dec. 19, 2023

Accepted: Jan. 25, 2024

Published Online: Apr. 11, 2024

The Author Email: Weigang Hou (houwg@cqupt.edu.cn)

DOI:10.3788/AOS231949

CSTR:32393.14.AOS231949

Topics

laser devices and laser physics

Lasers and Laser Optics

Laser physics

laser manufacturing

Instrumentation, Measurement and Metrology