Fig. 2
From: Image processing with Optical matrix vector multipliers implemented for encoding and decoding tasks
Simulation results comparing the performance of Type 1 and Type 2 multipliers in decoding tasks. a Autoencoder architecture trained on the MNIST dataset with 784 input/output neurons and 16 encoding neurons. b Weight matrix handling both positive and negative elements for optimal decoding results. c, d Visual comparison of the fan-out input patterns, weight matrix patterns, post-multiplication patterns, and fan-in images (top panels for positive weights and bottom panels for negative weights), followed by subtraction between the pairs and sigmoid activation to reconstruct a digit “7” using Type 1 and Type 2 multipliers. e MSE loss comparison between digital, Type 1, and Type 2 multipliers under varying optical distortion levels. Top panels compare the corresponding reconstructed results. Optical distortion is quantified using vertical and lateral shifts (random displacements within ±X% of the image width, modeled as a uniform distribution) and shear transformations (angular distortions within ±X radians). See Supplementary Fig. S3 for schematic illustrations of these distortions
