Optical Wavefront Engineering and Optoelectronic Techniques for Neuromorphic Computing

Machine learning has emerged as a powerful tool for solving complex problems in various fields, including optics. The scalability of compute resources and energy consumption for larger AI models is, however, a major concern. Optics can be used to implement matrix-vector-multiplications, essential for neural networks, in an energy-efficient manner. However, optical non-linearities are difficult to implement with traditional optical methods. This work demonstrates an optoelectronic device that implements an optical MVM and an electronic non-linearity without the need for high intensities. The device operation is experimentally demonstrated and a scaled-up version is proposed that demonstrates an order of magnitude higher energy efficiency when compared to conventional hardware. In addition to improving ML implementations with optics, ML can be used to solve problems in optics, including in holography. Conventional holography uses a phase hologram to generate a target intensity pattern upon diffraction. A novel alternative is introduced: holography using only polarization. Conventional phase retrieval algorithms are insufficient to optimize such a hologram. This work demonstrates the use of gradient based optimization of neural networks incorporating a differentiable numerical model of polarized light propagation to optimize for a target intensity distribution as well as the joint optimization for a target intensity and polarization distribution post diffraction.