Artificial Intelligence (AI) in Contested Multi-Domain Coalition Operations

Watch Overview Video

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, there is a need for:

Adaptable AI

AI which can rapidly adapt in dynamic situations and learn as the operation proceeds exploiting synergies between humans and machine intelligence (inc. novel Neuro-Symbolic Learning (NSL) AI systems which combine reasoning and deep learning);

• Adversarial Domain Adaptation Learning for Accelerating Artificial Intelligence Based Military Solutions. Adjusting machine learning classifiers for new environments using limited training samples through generative adversarial networks (GANs)
• The FastLAS System for Interpretable Machine Learning. Logic-based AI that learns rules from examples.
• Adaptive Artificial Intelligence Systems for Human-Machine Federated Decision Making. Adapting neural layers using human provided or learned hard logic at the symbolic layer.
• Cogni-Sketch: Enabling Rapidly Formed Human-Agent Coalition Teams through Extensible Information Exchange. Software platform enabling human-agent interaction.
• Adapting Artificial Intelligence Systems to Recognise New Patterns of Distributed Activity. Improved human-AI teaming, AI learning (inc. NSL) and DAIS technology integration: running in real-time at edge.
• Advancing Artificial Intelligence with Neural-Symbolic Learning and its Application to Generative Policies in Distributed Coalition Operations. Adapting to changes between data used for training an AI and reality.
• Reinforcement Learning for Military Network Control. Efficient learning of complex decision spaces for real-time control of the network.
• State-Action Separable and Embedding for Reinforcement Learning. Taming complexity to enable learning of optimal network control policies.
• Characterizing New Devices on a Network with Zero-Shot Learning. Characterising previously unseen devices by inspection of their traffic.
• Online Multi-Task Learning with Long-Term Memory. Learning to recognize and adjust network analytics for different operating environments.
• Enhancing Situation Understanding through Negative-Ties Enhanced Pipelines. Improving AI analysis with user knowledge.

Trusted AI

Enable human users to rapidly achieve an appropriate degree of trust in AI systems when making high-stakes decisions;

• Achieving Rapid Trust of Adaptable Artificial Intelligence Systems. Conceptual framework for trust calibration; identifies 6 roles requiring different explanations.
• Uncertainty-Aware Artificial Intelligence and Machine Learning. Revealing when the AI does not know in real-time at the edge.
• Real-Time Explainable Artificial Intelligence: Time-Series and Multi-Modal Data. Revealing what the AI is paying attention to in real-time at the edge.
• Testing the Reliability and Consistency of Explanation Metrics. New tests expose problems with saliency metrics.

Resilient AI

AI which is resilient to adversary attacks which seek to deceive the AI systems;

• Model Poisoning Attacks and Defences in Federated Learning. (aka Distributed Coalition AI). Defending against partner attacks.
• Efficient Attacks Using Side-channels. Defending against adversary use of AI explanations to develop attacks.
• Gradient Free Attacks on Multiple Modalities. Reducing the amount of data required, about a model, to launch a successful deception attack.

Distributed Coalition AI

AI systems able to share data and models with partners whilst operating under a range of privacy constraints and in degraded communications environments;

• Adaptive Federated Learning in Resource Constrained Edge Computing Systems. Sharing parameters, not data, and synchronising models with a minimal number of messages.
• Compressed Model Updates for Efficient Federated Learning. Minimising size of messages for shared learning and model synchronisation.
• One Shot Federation for Coalition Model Sharing. Incremental and intermittent shared learning using a representation of the data distribution (not the data itself).
• Coresets via Multipronged Data Reduction. Significantly reducing comms usage by summarisation of training data.
• Coresets Learning via Distributed Clustering and Local Gradients. Multi-dimension data summarisation (reducing comms usage) which also enables machine learning.

Integrated Distributed Analytics

Able to integrate analytic services in (near) real-time with partners in degraded communication environments;

• Vector Symbolic Architectures and Hyperdimensional Computing for Coalition Operations - An Overview. Dynamic decentralised discovery of assets (e.g. information services & data) and chaining them together (i.e. workflow construction and orchestration) to perform a task.
• Dynamic Communications Replanning using a Vector Symbolic Architecture. Demonstration of maintaining network connectivity via the vector symbolic architecture technology.
• Combining Vector Symbolic Architecture Aspects and Artificial Intelligence Services Using Edge Deployment. Feasibly integrating centralised control and distributed adaptability of coalition services in tactical environments.
• Semantic Vector Mapping for Coalition Operations. Achieving service interoperability without having to use an agreed set of terms to define the service.
• A Compositional Reinforcement Learning Framework for Workflow Generation. Learning how to construct coalition workflows with sparse rewards by leveraging the inherent hierarchical structure present in the application domain.

Edge AI

AI able to operate on the constrained computing environment at the edge of tactical networks;

• Edge AI Software Development Kit for Coalition Analytics. Enabling algorithm development, test and management.
• Model Pruning for Efficient Federated Learning in Coalitions. Modifying AI models so they can run on edge devices with minimal loss in performance.
• Energy Efficient Vector Symbolic Architectures (VSA) using ‘In Memory’ Hyperdimensonal Computing. Significant energy saving using novel ‘in memory’ computing hardware.
• Energy Efficient Vector Symbolic Architecture Using Spiking Neural Networks. Potential of sparse VSA to enable significant energy saving using novel SNN computing hardware.
• Leveraging Binarised Neural Networks for SDC Control. Using binary representations of model weights to allow Machine Learning models on mobile hand-held devices to make inferences.