Passive nonlinear activation and Laguerre-Gaussian modes in a convolutional femto-optical neural network

We present an optical framework that integrates a nonlinear activation function into the first layer of a convolutional neural network (CNN) using a 4f optical system. A femtosecond-pulsed 1550 nm laser is modulated by trainable phase weights using a spatial light modulator (SLM) in the Fourier plane. A passive nonlinear operation, two-photon absorption (TPA), is induced through beam compression before detection by a CMOS camera operating outside the laser’s fundamental wavelength. Our system classifies MNIST digit classes using nonlinear optical signals and is compared to a visible-range linear optical CNN. Performance is evaluated through confusion matrices, highlighting the potential of TPA-based all-optical CNN layers for low-power, high-speed image classification