Leveraging Lithium Niobate on Insulator Technology for Photonic Analog Computing

Machine learning has experienced an unprecedented growth over the last decade with graphics processing units being instrumental to this success, used as accelerators for the computations required during training and inference in artificial neural networks. However, the increasing complexity of the employed neuromorphic models might pose issues in terms of required computing resources and energy to run these tasks. Optical solutions, especially exploiting integrated photonics, are promising to effectively run neuromorphic tasks, exploiting highspeed and low-power elements. In this paper, we propose to exploit cascaded low-loss and low-driving-voltage travelling wave Lithium Niobate on Insulator (LNOI) modulators to perform multiply-accumulate operations at high speed and low power consumption. Thanks to the moderate losses, the same input can be split to multiple weight modulators, which increases the energy efficiency. Simulations of computer vision tasks on well known datasets show that the proposed solution, notwithstanding the limitations due to the analog photonic physical layer, achieves a good accuracy with limited degradation with respect to traditional digital electronic implementations.