Machine learning (ML) has achieved phenomenal success in revolutionizing a number of science and engineering disciplines over the last decade. Naturally, it is also being reckoned as a powerful technology to transform future fiber-optic communication systems. Over the past few years, we have seen extensive research efforts by both academia and industry to assimilate and gain from ML paradigm in several aspects of optical communications and networking. However, despite years of rigorous research, ML has not yet gained broad acceptance in commercial fiber-optic networks. In this article, we identify major common factors which are currently hindering widespread adoption of ML in practical optical networks. As ML-based methods are inherently data driven, we particularly highlight critical data-related issues as well as intrinsic limitations of the ML algorithms. Taking two important use-cases, i.e., quality-of-transmission estimation, and proactive fault detection and management as examples, we elucidate how these limiting factors shrink the deployment prospects of ML-based solutions. We also briefly discuss main challenges faced by ML-assisted methods in seven other key areas of fiber-optic communications. Finally, we suggest some useful strategies that can help alleviate existing obstacles, thus paving the way for vast deployment of ML -powered tools in real optical network infrastructures.
