CNN algorithms have been pushed to the edge in a race to decrease training time and achieve higher accuracy scores. The cost of implementing these state-of-the-art architectures includes the necessity of using powerful processors, which however, are limited by clock frequency, memory access time, and mainly, power consumption. Although integrated photonic implementations of CNN accelerators have the potential to alleviate these issues, the lack of scalable on-chip optical nonlinearity and the relatively high loss of photonic devices can limit the performance of such Photonic CNNs (PCNNs) in more complex and large-scale tasks. Nevertheless, the utilization of Lithium Niobate on Insulator (LNOI) in photonic platforms may be a solution to overcome these obstacles, paving the way for the implementation of large-scale PCNNs.
The aim of this project is to simulate, fabricate, and benchmark a high-speed, low-latency, large-scale Photonic CNN integrated on-chip, based on Lithium Niobate on Insulator platforms. Systems will be fabricated at RMIT, and simulations will be carried out at INL/RMIT as accelerators for large-scale image processing, with performance projected to optimal technology parameters in order to achieve sub-nanosecond image classification.
Watch a video about Mateus’ project:
