Enhanced photonic accelerator design for efficient vector-matrix multiplication in deep learning

Matrix multiplication with weights is a fundamental operation in deep neural networks, crucial for tasks reliant on data-driven modeling. However, computational limitations often hinder these operations. To address this, we propose photonics neural networks, offering faster processing speeds compared to electronic systems. Our paper aims to develop two distinct photonic neural networks. We use micror-ing resonators as neurons, exploiting their varied transmission properties for weight updates. Our experiments achieved a maximum transmission drop of approximately 0.7 dBm and a maximum spectral shift of around 1 nm per voltage change. Furthermore, we propose and explore two methods for implementing summation and non-linear activation using photonics devices with various materials, investigating the design constraints associated with each approach. Our experimental results corroborate our simulation predictions, validating the feasibility of our designs.