Edge AI Software Development Kit for Coalition Analytics
Military / Coalition Issue
Typically, AI involves large amounts of data being collected, transferred and processed in data centres. This presents severe challenges to the use of AI at the edge of the network due to the limited connectivity and compute capability available; to which there are added challenges associated with trust, confidentiality and integrity in coalition operations.
Core idea and key achievements
Developed an initial Edge AI software development kit (SDK) to enable production and test of tailored solutions for AI in these situations. The SDK offers flexibility to cover different usage patterns, such as one or a combination of: efficient data collection; data anonymisation; edge training; edge inferencing; edge model adaptation; collaborative training and collaborative inference.
Implications for Defence
Edge AI allows the military to deploy, train and run more effective AI systems at the edge. This brings a clear advance in capability with two major advantages being: lower bandwidth requirements; and better AI capability at the edge of the network. An example pattern for the military would be: from a catalogue of pre- trained AI models held in central storage, the best model for the task (e.g. vehicle identification) is selected through a ranking process; this model is pruned to compress it where it can then be efficiently transferred to edge devices (because it is smaller); the pruned model is executed on the edge devices and due to pruning, more effective models can be run on an equivalent device (makes better use of bandwidth and CPU/memory); as the situation progresses the edge devices co-operate to tweak the model using transfer learning (one device can help another) and federated learning (devices work collaboratively for mutual benefit) at the edge; federated inference (devices jointly make predictions) may also take place; finally the model can optionally be transferred efficiently (because it is compressed) back to the central catalogue if desired such that it can be selected and re-used in future operations.
Readiness & alternative Defence uses
The capability for version 1 of an Edge AI SDK that incorporates all the described functionality is currently in progress within IBM Research. It incorporates and extends DAIS research such as coresets, federated learning and model selection. It also incorporates some non-DAIS IBM research and integrates with existing edge deployment and management software. Further progress has been set out to provide a vision for the Edge AI SDK in the near future and a road map towards version 2. Various parts of the capabilities of the SDK are at different maturity levels that approximate to TRL levels 3, 4 and 5. Thus, ready to prototype a military Edge AI production facility, and experiment with a range of potential usage patterns.
Resources and references
Keith Grueneberg, Xiping Wang, and Seraphin Calo. “DAIS Edge Transition” Technical Report and Demonstration April 2022
Organisations
IBM UK, IBM US