Notice

Challenge 2: additional information and innovation challenge elements

Updated 10 August 2017

1. Challenge 2 additional information and innovation challenge elements

1.1 Free up personnel through the application of innovative use of machine learning algorithms and artificial intelligence (AI) for military advantage

We want to develop a greater understanding of all available data sources by leveraging AI technologies, and be ready to effectively exploit the increased data volume and variety that is expected in the future. This could be from many different sources, including sensor data, intelligence reporting, and open source media feeds. AI technologies have already demonstrated benefits in other domains for discrete information processing tasks such as image classification and face recognition, displaying better than human accuracy in some cases.

We would like to explore the application of such technologies in new Defence areas such as the automated identification and assessment of events in near-real time, to enable autonomous triage and alerting capabilities for commanders in order to flag significant incidents. Currently significant manpower is expended to manually process and analyse data, which can delay effective decision making.

1.2 Example within a military scenario:

Data is received from multiple inputs including: sensors (imagery, video, radar, sonar), intelligence reports, information from allies, open source media, platform systems (for example ship engine monitoring), and environmental parameters (such as weather). The system automatically triages, enriches and processes the integrated data volume, in order to extract patterns and anomalies, and to flag information which is pertinent to the operation. Data deemed important is immediately highlighted to the commander, and over time the system learns from interaction with key decision makers – what information is relevant? Based on historic patterns, the system predicts likely future scenarios for a variety of factors which are relevant to the operation, in order to optimise recommendations made to the decision maker.

2. What we are interested in specific to challenge 2

• solutions for component areas within an overarching information processing architecture that conform to open standards, solving aspects of the problems really well

3. What we are not interested in specific to challenge 2

• machine learning based solutions which are highly optimised for the input of training data, leading to problems associated with over-fitting and failure when environmental parameters change

4. Challenge 2 technical areas of investigation: free up personnel through the application of innovative use of machine learning algorithms and artificial intelligence (AI) for military advantage

This challenge seeks original and innovative solutions to free up operator time through the automation of currently manual information processing and exploitation tasks. Since 2012 the explosion of machine learning capabilities, principally driven by the commercial sector, has demonstrated that near-human level accuracy can be attained for data processing tasks such as face recognition, image classification, and automatic lip reading. MOD would like to exploit the underpinning technologies which drive these capabilities, and apply them in a Defence scenario such as automatically detecting and classifying the activities of assets as friendly, neutral, or hostile.

5. Specific areas of interest for challenge 2

5.1 Automated activity classification

MOD requires methods for automated detection and classification of activities and intents from multiple sensor types using state-of-the-art machine learning and AI. MOD is particularly interested in innovations which:

• deliver beyond simple object and feature extraction, and aim to develop methods for scene classification

• operate both in real-time at the ‘sensor-end’ and batch processing mode in a large data centre, exploiting the latest hardware developments

• require little or no training overhead and can be rapidly applied against new areas, potentially utilising semi-supervised and un-supervised machine learning methods

Such automated processing would reduce the manual burden on analysts, but would also reduce pressure on communication bearers, by automatically flagging the data of interest to operators.

5.2 Cognitive computing

MOD would like to exploit the advancements made in other sectors in automated voice recognition, natural language question answering (underpinned by advanced machine learning), and knowledge graphs, and to understand how these can be combined as digital assistants in support of military personnel in the future. We want to augment the human, and free their time through automation of manual tasks, rather than completely replace them. We require solutions to automatically flag adversary activity of interest, based on known knowns, but also identify previously unknown information of relevance. Another area of assistance could include the automated identification of incorrect/fake information which has been generated through digital deception, by cross-referencing across all information stores.

5.3 Combined human/machine derived models

MOD is interested in the combination of human-derived models, where the analyst makes use of a small amount of data and a-prior knowledge to generate the model; with machine-derived models, which generally require larger volumes of data and are largely driven by machine learning technologies. We would like to combine the best of both approaches in order to deliver an enhanced capability for defence.

Specific research questions include:

• how do we combine data and human derived models
• how do we build more robust statistical models of subjective measures (for example assessment of threat)
• how do we ensure data-driven models are transparent and understandable for analysts and operators?

5.4 Predictive analytics

We want to apply machine learning in support of predictive modelling to guide military decision making. MOD requires solutions which go beyond enhancing military understanding of current situations, but predicts future outcomes, including actions, anomalies, intent and movements, to guide decision makers in support of operational planning.