The FastLAS System for Interpretable Machine Learning
Military / Coalition Issue
When deploying artificial intelligence systems in military settings, it is often important that the behaviour of such systems can be verified, audited and explained. It is difficult to extract meaningful explanations from a black-box machine learning system, presenting a potential barrier to the adoption of machine learning in military settings. FastLAS is a new logic-based machine learning system that learns interpretable rules, making it possible to generate explanations for each decision.
Core idea and key achievements
Previous state-of-the-art logic-based machine learning algorithms could be grouped into two categories: approximate and exact systems. The former do not guarantee finding the best set of rules and tend to perform poorer when evaluated on test data; the latter are systems that guarantee finding the best set of rules, but limited in their scalability over large search spaces. FastLAS is a new exact system that does scale to large search spaces.
Rather than optimising over the full search space straight away, FastLAS takes a novel approach of using the labelled data to carefully construct a small subset of the search space that is guaranteed to contain at least one optimal solution. It has been shown to be much more scalable, and 2-3 order of magnitude faster, than other current state-of-the-art exact systems, and able to learn rules that have a higher predictive accuracy than those found by state-of-the-art approximate systems. The first version of FastLAS has strong restrictions on the rules that can learn, limiting its applicability. The recently developed second version, called FastNonOPL overcomes many of these restrictions, widening the applicability of the approach; for example, FastNonOPL is capable of learning knowledge that is only indirectly observable, and can handle non-deterministic knowledge and data which are only partially defined.
Implications for Defence
Military operators will be able to deploy this system to tabular data to learn set of rules for prediction and decision-making, which can be expressed to humans in natural language. Existing Deep Learning models can be integrated with our logic-based Machine Learning systems to support learning of interpretable complex rules from unstructured data in the presence of limited data . The systems can also take into account expert domain knowledge from different parties when learning in federating setting. Our system can generate explanations to help decision makers understand its output and increase trust in their outputs.
Readiness & alternative Defence uses
Our systems are at TRL 4 level. They have been validated in the lab and are ready to be deployed in a real environment. FastLAS has also been used in several other DAIS achievements (e.g., 1c.02, 1c.07, 1c.16). Although it has mainly been used to learn policies within DAIS, FastLAS is a general-purpose rule-learner that can be used in a wide variety of applications.
Resources and references
- Law, Mark, et al. “FastLAS: scalable inductive logic programming incorporating domain-specific optimisation criteria.” Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 34. No. 03. 2020.
- IJCAI 2021 (To appear): Scalable Non-observational Predicate Learning in ASP. Law, M., Russo, A., Bertino, E., & Broda, K.
- GitHub repository: https://github.com/spike-imperial/FastLAS
Imperial College London, Purdue University, Universitat Pompeu Fabra.