Enhancing Neuromorphic Photonic Hardware Performance through Neural Architecture Search

Current neuromorphic photonic neural networks cannot achieve the performance of electronic neural networks due to the presence of physical constraints, such as noise and distortions, affecting the implementations based on analog photonic hardware. This paper proposes the exploitation of Neural Architecture Search (NAS) tailored for Photonic-Aware Neural Networks (PANNs), a class of neural networks amenable to a neuromorphic hardware implementation. In this way we are able to obtain PANN architectures that not only meet these photonic constraints but also to outperform existing photonic models. Experimental results on the CIFAR-10 dataset demonstrate that exploiting NAS while addressing photonic constraints can signifi cantly improve the performance of PANNs, obtaining results that are comparable with state-of-the-art electronic networks. Indeed, the best-performing configuration achieved an accuracy of 95%, a performance similar to electronic counterparts while complying with photonic constraints.