Adaptive Artificial Intelligence Systems for Human-Machine Federated Decision Making
Military / Coalition Issue
Military operations typically involve multiple parties and a variety of devices that collaboratively adapt and respond to rapidly evolving situations. To achieve operational goals faster and with better outcomes military information processing systems must be able to support human-machine federated decision making, that use and share insights learned from (local) data by partners, adapt them to contextual changes, and integrate prior human knowledge.
Core idea and key achievements
Limited availability of data (due to changing contexts) can be compensated by domain-specific semantics and constraints. Pure Deep Learning (DL) methods are effective in extracting insights from raw data (e.g. sensors) but require large quantities of data with distributions similar to that used during training. We have developed techniques for injecting semantics and constraints into Deep Learning AI systems to allow data-efficient domain adaptation of predictive models and coherent federated inference for decision making. Specifically, our techniques enable:
- representation of semantic constraints and human knowledge in an interpretable program matrix form, and exact symbolic inference to be performed in a pure differentiable way.
- end-to-end integration of differentiable inference with DL-based systems to control and guide their adaptation, during the learning process, to new shared data, boost their predicted accuracy in the presence of limited data, and provide explanations to predicted inference from data.
Our techniques are able to learn predictive models from heterogeneous data in a more accurate and data efficient manner than pure DL methods taking into account partial information, and spatial and temporal constraints.
Implications for Defence
Military operators will be able to enhance DL-based systems using expert domain knowledge and inject constraints for federating models from different parties. Existing DL models can be adapted quickly to changing conditions (e.g., different environments, changes in weather, lighting condition), than a pure DL method. Our system can generate explanations to help decision makers understand its output and increase trust in AI models.
Readiness & alternative Defence uses
TRL 3/4. The system will be made available as open source software. While the approach has been tested on images and structured data, it is applicable in general to other time-series data (e.g. video, network traffic data).
Resources and references
- Aspis, Yaniv, et al. “Stable and supported semantics in continuous vector spaces.” Proceedings of the International Conference on Principles of Knowledge Representation and Reasoning. Vol. 17. No. 1. 2020.
Imperial College London, IBM UK.