Figure 10
(a) Schematic of the proposed CNN architecture using IM-MRMs for modulation and weight banks. The kernel dimension is 3 \(\times \) 3, and the number of input channels is eight. Nine lasers with different wavelengths are used to feed the system. Input images are encoded in transmitted intensities of all-pass IM-MRMs by modulation banks. Kernel values are loaded to add-drop IM-MRMs in weight banks, which then perform the dot products with input signals. Finally, all signals are accumulated at the output port, resulting in the convolved feature. Activation functions, pooling, and fully connected layers are followed offline. (b) Transmission spectra of nine cascaded add-drop IM-MRMs at through (blue) and drop (red) ports with the correlated mapping value of 1 (dashed curves) and 0.5 (solid curves) in each wavelength channel. The wavelength channel spacing is limited with \(\delta \omega \) = 0.5. (c) Attainable mapping values for nine cascaded IM-MRMs with the limited wavelength channel spacing. A common range of [–0.5, 0.5] is used for the proposed CNN system. (d) Performance of the proposed CNN system on MNIST recognition task versus input precision.
