Table 1 Performance comparison of SmartLight OSS-CNN and a respective digital CNN architecture

From: Photonic neuromorphic accelerator for convolutional neural networks based on an integrated reconfigurable mesh

Scheme

Number of FCLs (units per layer)

MNIST accuracy (%)

\({\bar{{{\rm{P}}}}}_{{{\rm{GPU}}}}\) (W)

Single-layer CNN

1 (10)

98.6

50

2 (117,10)

99.1

50

7-node ideal

OSS-CNN

1 (10)

98

33.3

2 (111,10)

98.6

34.3

3-node experimental OSS-CNN

1 (10)

95.9

32.8

2 (80,10)

97.7

34.3

  1. Comparison of the performance of the optimum ideal and experimental SmartLight optical spectrum slicing convolutional neural network (OSS-CNN) with a respective digital single-layer CNN. The table details the number of fully-connected layers (FCLs) in the back-end of each scheme, the number of units per layer, the testing accuracy on the MNIST task, and the mean power consumption of the graphics processing unit (\({\bar{{{\boldsymbol{P}}}}}_{{{\boldsymbol{GPU}}}}\)) during training.
Back to article page