Scalable optical learning operator

Today's heavy machine learning tasks are fueled by large datasets. Computing is performed with power hungry processors whose performance ultimately limited the data transfer to and from memory. Optics one of powerful means communicating processing information there intense current interest in optical for realizing high-speed computations. Here we present experimentally demonstrate an computing framework based on spatiotemporal effects multimode fibers a range classifying COVID-19 X-ray lung images speech recognition predicting age face images. The presented overcomes energy scaling problem existing systems without compromising speed. We leveraged simultaneous, linear, nonlinear interaction spatial modes as computation engine. numerically showed ability method execute several different accuracy comparable digital implementation.