By Casimer DeCusatis
Machine learning has received a lot of attention recently, as it finds applications in everything from self-driving cars and digital personal assistants to language translators and fraud detection systems. It’s apparent that while machine learning (and its cousin, artificial intelligence) can’t solve every problem, there are a set of well-defined issues that are very well suited to this approach. There’s even a special symposium planned at the next OFC meeting, led by researchers from MIT and the University of California, Davis, to consider machine learning in photonic systems. In this blog, we’ll consider what makes a good candidate problem for machine learning, and look at some of the emerging solutions in optical communication physical layer and transport networks.
Is machine learning a good solution?
As discussed at a recent NIST workshop, there are several relevant criteria which help determine whether machine learning is a good solution for a given problem. First, some problems lack a good physics-based model due to insufficient understanding of the fundamental issues involved (i.e. model-deficit) or because a model exists but the algorithms required are too complex or require too long execution times (i.e. algorithm-deficient). Next, since many machine learning systems struggle to explain how they arrived at a given solution, a good problem is one for which the details of how the task is completed are not relevant. This is expected to be an issue for many network service providers, whose existing infrastructure and business process relies on a much more deterministic approach. Finally, we can only apply machine learning if a sufficiently large, diverse, and representative labelled set of training data exists or can be simulated. In fact, since machine learning relies so heavily on good training data, the cost and availability of this data is a significant factor in determining whether or not to apply machine learning.
Machine learning in optical network transport
There are some key problems facing optical network transport that are a good match for machine learning solutions. Transport network operators are always seeking new ways to increase capacity while lowering capital and operating expenses. Enabling these solutions might include networks that can self-configure based on real time physical layer data. This includes adaptively changing the routing, spectrum, or modulation format in response to changing load conditions or to correct a network failure. Transport layers might also self-optimize to avoid congestion or improve other performance metrics. Conventional systems deal with these issues through a combination of over-provisioning and safety margins designed into the network, but these approaches waste valuable network resources. Machine learning might allow networks to maintain or improve performance without adding capacity, through a more efficient utilization of existing components.
Solutions and optical network technologies
The realization of such solutions depends on a number of underlying optical network technologies. As noted earlier, availability of good quality data sets is essential for a successful deployment of machine learning. Data collection, storage, and processing all need to be enhanced if existing networks are going to exploit machine learning solutions. For example, machine learning can compensate for nonlinear transmission issues, provided that a received data set of performance parameters is available. Flexible coherent line interfaces can help provide the required data on which machine learning optimization can be based. Further, software defined networks (SDN) with centralized control planes provide a mechanism for adaptively adjusting network performance in real time, based on the results of machine learning algorithms.
Near term examples of machine learning in photonic systems
Some near term examples of machine learning in photonic systems include modeling key devices, such as EDFAs or ROADMs in long distance networks. Optical performance monitoring is also being investigated, which requires capturing end-to-end behavioral data on quality of service. Other efforts model physical layer impairments (such as bit error rate, eye diagrams, or optical signal-to-noise ratios) and even attempt to predict traffic flow patterns and proactively compensate for performance issues. Modulation formats may be recognized and adjusted using amplitude histograms or Stokes space parameters.
Machine learning and lifecycle management
But near-term examples are not the whole story. Practical solutions will be able to guarantee lifecycle management for a machine learning solution. In particular, sustaining fairness in the model can require additional algorithms to detect and remove bias. Models need to be robust in the face of potential data contamination or tampering, providing an interesting new set of cybersecurity challenges. Validating, pre-processing, or cleaning data sets are likely necessary steps, but also raise questions about how data sets should be updated and released, as well as who should have access to these data sets in the industry and research communities. Despite these concerns, it’s widely recognized that machine learning can provide a huge efficiency boost to optical communication networks and you won’t want to miss all the latest work in this area at this year’s OFC meeting.
Do you have a photonics problem that seems well suited to machine learning? Drop me a line on Twitter (@Dr_Casimer) and maybe we’ll discuss it in a future blog.
Check out the content of the symposium, The Role of Machine Learning in Optical Systems and the Role of Optics in Machine Learning Systems, to be held at OFC 2021.
Posted: 8 February 2021 by
Casimer DeCusatis
| with 0 comments