Harnessing Optoelectronic Noises in a Hybrid Photonic Generative Adversarial Network (GAN)

Abstract Integrated programmable optoelectronics is emerging as a promising platform of neural network accelerator, which affords efficient in-memory computing and high bandwidth interconnectivity. The analog nature of optical computing and the inherent optoelectronic noises, however, make the systems error-prone in practical implementations such as classification by discriminative neural networks. It is thus imperative to devise strategies to mitigate and, if possible, harness optical and electrical noises in optoelectronic computing systems. Here, we demonstrate a prototypical hybrid photonic generative adversarial network (GAN) that generates handwritten numbers using an optoelectronic core consisting of an array of programable phase-change optical memory cells. We harness optoelectronic noises in the hybrid photonic GAN by realizing an optoelectronic random number generator derived from the amplified spontaneous emission noise, applying noise-aware training by injecting additional noise to the network, and implementing the trained network with resilience to hardware non-idealities. Surprisingly, the hybrid photonic GAN with hardware noises and inaccuracies can generate images of even higher quality than the noiseless software baseline. Our results suggest the resilience and potential of more complex hybrid generative networks based on large-scale, non-ideal optoelectronic hardware. The demonstrated GAN architecture and the proposed noise-aware training approach are generic and thus applicable to various types of optoelectronic neuromorphic computing hardware.