AI boosts photonics and vice versa

With the advent of digital technology, big data, and deep learning, the last decade has undoubtedly seen a dramatic change in the landscape of science, industry, defense, and medicine.^1–4 Although the notion of artificial intelligence (AI) was first introduced in 1955, the breathtaking evolution of deep learning in the last several years has shifted the field of AI from largely theoretical studies in the discipline of computer science to the development of real-life applications such as astronomical observation,^5,6 drug design,^7,8 medical imaging,^9,10 flow cytometry,^11–14 distributed sensing,¹⁵ speech recognition,^16,17 finance,^18,19 and advertising.^20,21 In general, AI simulates human intelligence to identify recurring patterns based on past events and develop predictive models with the help of machine learning algorithms that recognize trends and predict future events. According to recent reports, AI-generated results in these applications have been proven comparable or even superior to the performance of human experts.¹⁰ This major success stems from deep learning based on techniques such as artificial neural networks and deep neural networks which are capable of revealing obscured patterns beyond human recognition.^1,2

There is an excellent synergy between photonics and AI (Fig. 1). Photonics boosts AI since it facilitates the laborious process of data collection which is an essential part of AI development. Also, AI boosts photonics since it enhances the capability of optical technology (e.g., optical sensing, optical imaging, optical diagnosis, and optical communication). This synergistic effect is indicated by the rapidly increasing number of AI-related publications in photonics journals in the last several years (Fig. 2). Specifically, autonomous cars use an array of high-speed optical sensors to acquire environmental information and use it to make decisions based on machine intelligence in real time.²² Another example is deep learning of numerous optical microscope, x-ray, and computed tomography images of cells and tissues for AI-assisted diagnosis of various diseases and disorders such as Alzheimer’s disease, Parkinson’s disease, skin cancers, gastrointestinal diseases, tuberculosis, atherothrombosis, and COVID-19.^23–28 In optical communication, AI-based techniques such as the characterization of optical network components, performance monitoring, and the mitigation of optical nonlinearities have been developed to predict and optimize the performance of optical communication systems and networks.^29,30 In astronomy, deep learning has been applied to optical telescope images of strong gravitational lensing to identify dark matter and hence to map the vast undiscovered content of the universe.³¹ Recently, researchers have started to construct and implement functional AI algorithms for photonics-based quantum computing.^32,33 In the industrial sector, Sony has already launched an image sensor equipped with AI image analysis and processing functionality on its logic chip, called an intelligent vision sensor.³⁴ 

Driven by the alliance between photonics and AI, the objective of this Special Topic is to highlight basic principles, advanced techniques, and applications of photonics with AI. Specifically, the Special Topic contains a collection of 12 original research articles and 2 perspective articles contributed by theorists, modelers, and experimentalists. Luo et al.³⁵ report a genetic neural network as a hybrid algorithm to reinforce light focusing in disordered media such as biological tissues, together with experimental results that demonstrate improved wavefront shaping performance. Yan et al.³⁶ introduce optofluidic time-stretch quantitative phase imaging without phase measurements by taking advantage of virtual phase images constructed using a deep neural network trained with numerous pairs of bright-field and phase images and demonstrate its application to accurate classification of leukemia cells and white blood cells. Teğin et al.³⁷ present a method based on deep neural networks to analyze and control spatiotemporal nonlinear interactions in multimode optical fibers (e.g., cascaded stimulated Raman scattering and supercontinuum generation). He et al.³⁸ show high-resolution incoherent X-ray ghost imaging with a single-pixel or bucket detector based on a multi-level wavelet convolutional neural network with a compressed set of Hadamard matrices incorporated into it. Giacoumidis et al.³⁹ demonstrate a real-time fiber nonlinearity compensator based on a field-programmable gate array with a machine learning algorithm for energy-efficient coherent optical networks. Montresor et al.⁴⁰ report a deep-learning-based method for de-noising phase data (i.e., speckle noise in phase measurements) in digital holographic interferometry and use it to experimentally demonstrate improved de-nosing performance in wide-field digital holographic vibrometry. Qiao et al.⁴¹ present their development and application of deep learning for video compressive sensing within the scope of snapshot compressive imaging and show enhanced performance in speed, accuracy, and flexibility. Huang et al.⁴² propose and demonstrate a simple strategy for continuous multichannel control of weight banks in microring resonators with excellent accuracy and precision and discuss its utility to large-scale photonic integrated circuits including photonic neural networks. Zhang et al.⁴³ show label-free colorectal cancer screening with a high classification accuracy of 99% based on spatial light interference microscopy assisted by deep learning in order to economize the clinician’s time and effort in staining of thin tissue slices in pathology. Valensise et al.⁴⁴ report a deep-learning-based technique to remove non-resonant background from broadband coherent anti-Stoked Raman scattering spectra, a spurious signal that contaminates the spectra. Qian et al.⁴⁵ demonstrate a method based on deep learning for unifying phase retrieval, geometric constraints, and phase unwrapping into a comprehensive framework for absolute three-dimensional shape measurements. Bosworth et al.⁴⁶ investigate machine learning attacks against a new class of nonlinear silicon photonic cryptographic devices known as physical unclonable functions that leverage nonlinear optical interactions in chaotic silicon microcavities. Lugnan et al.⁴⁷ review photonic neuromorphic computing with a strong focus on photonic reservoir computing, an area of photonic neuromorphic computing powered by the rise of deep learning. Applegate et al.⁴⁸ review recent advances in high-speed diffuse optical imaging, including deep learning, for a diverse range of biomedical applications.

In conclusion, this Special Topic is intended to highlight the recent advances of the synergy between photonics and AI. It is our hope that the Special Topic will serve as a forum for students and young researchers to further extend the potential of AI-boosted photonics or photonics-boosted AI. We are grateful to Editor-in-Chief Benjamin Eggleton and Journal Manager Erinn Brigham for the technical assistance with publishing.