Optical Convolutional Neural Network with Atomic Nonlinearity

Abstract

In the past decade, machine learning techniques, in particular artificial neural networks (ANNs), have been widely introduced in industrial applications and have played a more significant role in fundamental research. However, electronically implemented ANNs incur huge computational costs. In contrast to electrons, photons enable massive and parallel interconnections with high computational efficiency. Here we demonstrate an optical convolutional neural network in which linear operations are implemented by lenses and spatial light modulators (SLMs), while an optical nonlinearity is realized in the form of a cesium vapor cell as a saturable absorber. We use the handwritten digit dataset MNIST [1] to train and benchmark the optical convolutional neural network (OCNN). In our experiment the convolution is performed by pointwise multiplications in the Fourier plane, based on the convolution theorem. The digital micromirror device (DMD) of SLM selectively reflects the laser beam, thus the image is encoded in the spatial intensity distribution of the laser. By displaying and reflecting the pattern or complementary pattern of a circle on the SLM in Fourier space, a high-pass or low-pass filter that selects or deselects edge characteristics is realized. Moreover, we simulate the optical system, train the CNN and extract a two-dimensional kernel pattern from our simulation which is inserted into the optical setup. Using two lenses and a second SLM, we can manipulate the Fourier transform of the image to convolve the input image and the kernel. Nonlinear activation functions are realized optically as well. In this work we use a cesium atomic vapor cell for this purpose. An 894 nm laser is used to excite the cesium D1-transition in the vapor cell. The excited state population settles to the steady state when the atoms become saturated, thus a nonlinear relationship between the input power and output power is obtained. The scheme presented in this thesis provides a strategy for an energy efficient alloptical neural networks to be developed in the future.