Adapting Artificial Intelligence Systems to Recognise New Patterns of Distributed Activity
Military / Coalition Issue
Military operations typically involve working with partners to resolve rapidly evolving situations where adversaries are adapting their tactics, techniques and procedures, and the behaviour of the civilian population is changing. Thus, military information processing systems need to be able to recognise significant patterns of activity which are distributed in time and space in near real-time, without generating too many false alarms.
Core Idea and Key Achievements
In the language of information processing, this requires the ability to recognise the relationship between a set of individual events. Deep learning Artificial Intelligence (AI) systems are constantly improving their ability to recognise such individual events. We have developed techniques which enable:
- Humans to inject new knowledge, or hypotheses, about patterns of activity through addition of new rules (which means patterns can be recognised in situations where there is insufficient time or data to train a deep learning model);
- Significantly-reduced need for training data (especially important when examples of the patterns-of-interest are relatively rare), faster AI model training, and improved detection accuracy over ‘pure’ deep learning approaches.(1)
Further we have demonstrated that such architectures are loosely coupled (so suitable for open architectures) and the resulting applications can run on edge of network devices (so are suitable for tactical sensor systems).
(1) In trials our techniques trained the best-performing model 3-4 times faster and with up to 100 times reduction in annotated training data required.
An example AI system configuration with deep learning-based services (VGGish & AudioNN) and then by a probabilistic logic program (DeepProbCEP) with learned probabilities.
Implications for Defence
Enables Defence to process in near-real time streaming data from a future Internet of Battlefield Things (including network, CEMA (cyber electromagnetic activities), ISTAR (intelligence, surveillance, target acquisition, and reconnaissance) and HUMS (health and usage monitoring systems) data). Enables Defence to adopt AI systems which learn and adapt at the ‘pace of the fight’ – necessary to ensure our tempo of understanding and action is not over-matched by our adversaries.
Readiness and Alternative Defence Uses
This work is technology readiness level (TRL) 3/4. The individual components are available as open source software – DeepCEP and LiveEvents. While the approach has been demonstrated on video and audio data, it is applicable in general to other time-series data (e.g., electro-magnetic spectra and cyber traffic).
Resources and References
- Vilamala, M. R., et al. “A hybrid neuro-symbolic approach for complex event processing in noisy and adversarial settings.” Proc 36th International Conference on Logic Programming. 2020.
- Xing, Tianwei, et al. “Neuroplex: learning to detect complex events in sensor networks through knowledge injection.” Proceedings of the 18th Conference on Embedded Networked Sensor Systems. 2020.
Cardiff University, UCLA, ARL