Fig. 7
From: Image processing with Optical matrix vector multipliers implemented for encoding and decoding tasks
Analysis of MSE loss and energy efficiency across different SNR levels and EDs in the ONN-based autoencoder. a MSE trends for MNIST test data across varying SNR levels (bottom x-axis), with corresponding optical energy per pixel required for achieving MVM operations (top x-axis). The systems discussed here are trained with built-in noise tolerance mechanisms, incorporating on-system iterative tuning conducted with noise present, while models without such mechanisms are shown in Supplementary Fig. S8. b Total optical energy used for encoding and decoding tasks (red dots and left y-axis) and MSE loss, depending on the ED, estimated at an SNR of 20 dB. c Estimation of energy per MAC operation for encoding and decoding tasks, with and without considering multiplicative energy losses (94% total, ~6% efficiency) from electrical-to-optical conversion (30%) and transmission losses in components such as LCDs/SLMs (80%) and lenses (1%). This analysis focuses solely on input optical power and does not account for the additive operating power of active devices (e.g., OLEDs and LCDs)
