NSF Workshop on Measurements for Self-Driving Networks

ion are promising first attempts, more work will be required to develop the type of programming abstractions that are capable of unifying network monitoring, automated inference, and programmatic control. Moreover, such abstractions should be directly applicable to distributed settings. 5.2 Addressing the Trust Problem If AI/ML-based tools are used at all in some of today’s production networks, they typically entail simple learning models such as decision trees and SVMs. In fact, decision trees and their variants are currently one of the most commonly-encountered ML techniques in support for automated decision making in realworld production networks. They are in general preferred over the latest deep learning models, mainly because they are more lightweight, less complex, and more intuitive. While suitable for simple problems, their traditional use is usually ill-advised when the problem at hand is more complex. For example, using a decision tree model is not a recommended solution for a problem of practical interest for mobile providers— identifying rogue mobile devices using time series of various features collected from tens of millions of mobile devices. However, one of the main reasons why today’s network operators favor simpler learning models over more sophisticated and complex ones is that the latter are in general used as “black boxes”, unable to provide any insights or understanding. A keen desire for wanting to know when these black boxes succeed or fail (and why) and understanding why they produced a certain result and not something else is at the heart of the network operators’ demands for cracking these black boxes open and turning them into “white boxes” so that as users, they are able to understand, explain, and interpret their output; that is, trust them. Other areas where AI/ML is increasingly used and where failures can be catastrophic (e.g., autonomous vehicles, medical diagnosis, and legal decision making, etc.) also faced similar trust issues. Recognizing that the effectiveness of these systems is limited by the inherent ability of most of these learning algorithms to explain decision and actions to human experts, in 2017 DARPA launched a new program on “Explainable Artificial Intelligence (XAI)” [10]. The stated aims of XAI are “to create a suite of AI/ML techniques that (i) produce more explainable models, while maintaining a high level of learning performance (prediction accuracy), and (ii) enable human users to understand, appropriately trust, and effectively manage the emerging generation of artificially intelligent partners.” XAI is further expected to contribute to the development of “third-wave AI systems’, where learning algorithms “understand the context and environment in which they operate, and over time build underlying explanatory models that allow them to characterize real world phenomena.” Since the launch of the XAI program, we have witnessed exciting developments that hint at what may be possible in terms of explainable or interpretable AI/ML models. In particular, we can surmise why they promise to be a good match for the area of network automation where understanding how these advanced algorithm do what they do will be imperative for network operators/engineers and security analysts before they are willing to hand over consequential decision making to AI/ML-based tools and deploying them in their networks. To illustrate, a series of recent papers [2–4] describes new efforts for interpreting or explaining black-box models via model extraction. The basic idea is to approximate a complex black-box model using a more interpretable or explainable model. In particular, by considering decision trees as approximate models and assuming that the approximation quality or “fidelity” is high, any issues in the complex black-box model should be reflected in the approximate model (i.e., the extracted decision tree). Since decision trees are highly interpretable, users can now start explaining how the corresponding blackbox does what it does by examining the extracted high-fidelity decision tree instead. The feasibility of the idea behind explainable or interpretable AI/ML, including algorithms for extracting high-fidelity decision trees, has already been demonstrated in the context of black-box models such as random forests, DNNs, and some classical reinforcement learning models. In these documented cases, the constructed decision trees have been successfully used to debug and interpret the considered supervised learning models and to understand the control policies learned from traditional reinforcement learning models. However, despite these promising advances, there are many open problems and much remains to be done in this ongoing effort to transform AI/ML’s black-box models into “white boxes.” For example, decision trees quickly lose their interpretability or explanatory power as their depth increases and the number of paths through the tree becomes unwieldy. It will be important to explore alternative approximate models that match the high interpretability of decision trees and can be proven to also have high fidelity across a range of commonly-used black-box models. By being able to address these and similar issues (see for example [8]), it may be possible to persuade networking researchers to embrace explainable