Artificial Intelligence is changing the nature of contemporary military operations, from intelligence gathering and analysis to effectors at the tactical and strategic level, for the ADF, AI will serve as a powerful force multiplier explains Ben Whitham and David Liebowitz of Australian cyber company, Penten.
We know AI is a world-changing technology, but it does present some practical challenges before Australian Defence can realise its potential. We have found that timely, tangible and sustainable AI solutions can be implemented by integrating scientists into a multidisciplinary team and employing iterative and modular design that is not reliant on any single AI algorithm.
Challenges of AI Adoption in a Military Context
One of the most important factors in the successful implementation of new AI technologies, particularly those using Deep Learning, are the training sets which are used to teach the systems. These very large data sets (millions or even trillions of data points) are what allows AI systems to learn how to solve particular problems. Naturally this technology lends itself to repeatable experiences and context - you need to train how and where you are going to fight.
Here the military faces three challenges. First, commercial AI applications (trained on non-military data) may produce undesirable results when transitioned to a military context. Similarly, imported AI processes may require adaption to the Australian environment, utility and adversary, in a similar way that European cars may need to be adjusted for Australian road conditions. Finally, even after these factors have been taken into consideration, or sovereign AI data sets have been developed, military deployments rarely follow the same script as their predecessors. For example, terrain, adversary, platforms, tempo and force composition from East Timor were markedly different to those in the Persian Gulf Conflict.
AI algorithms may also require significant processing capacity and memory, and may take some time to execute. There is usually a constraint on one or more of those elements in military deployments. The challenge around processing capacity is further exacerbated by a preference for graphical processors to perform the calculations. Such processors, while common in high performance gaming machines, are not typically found at the scale desired in low-powered, small screen, military internet-of-things devices or in headquarters IT infrastructure.
AI systems are a very new technology and can sometimes produce incorrect or surprising results. These results might be misleading, or the outcome of inputs designed deliberately by an adversary to fool the AI. Military operators, even highly trained ones, are not machine learning experts, and are unlikely to be able to interpret such results or modify systems in the field.
Practical Approaches to Deliver Tangible Outcomes
A key approach to delivering practical results for the end user is to focus on solving specific problems as well as you need to. Avoid over engineering systems in terms of capacity, performance or computation, because these come at a cost and may impact the usability of the system, especially for land forces. We have found some classical models, which have been surpassed in many tasks by Deep Learning, can still produce useful results with limited training, at low computational cost and in shorter time frames.
If there is a need for more sophisticated models, consider adopting design redundancy into the AI system. We have seen favourable results when delivering systems, which implement multiple methods. These multiple methods may include less accurate algorithms, but produce results in a shorter period and/or with less information, processing capacity and memory. This could involve a solution that is able to draw on a portfolio of algorithms, selecting computationally simpler or faster models when memory, processing capacity or time is limited.
AI systems can be superseded quickly, so this approach also supports life-cycle upgrades when new methods are published. The key to success with a modular and multi-solution system approach is to focus on the model selection logic and clearly communicate the level of confidence in the output.
Solutions should also consider the training impact and the operational demands on the end users. Preference should be for systems that do not rely on the user being an expert. While sometimes advanced users might be present, the AI system should also ship with sensible default settings. If a training process or configuration is important, this should be simplified. Ideally, the AI engineering team should work closely with system engineers and human user interface designers. Together, they can ensure that the system provides immediate feedback on the consequences of configuration settings.
Another way of achieving practical outcomes in AI is to use an iterative design process with regular feedback from the end user community to refine the system. We have seen great value in these design cycles, particularly when we have matured the entire system rather than just the scientific or algorithm aspects. This approach requires a multidisciplinary team, which includes AI scientists and systems engineers working closely with end users (or an SME that understands the end user's needs and constraints).
The challenges of collecting large volumes of applicable data to train systems can be approached with the application of transfer learning: trying to use models trained for other tasks as a starting point. We have learnt to come prepared with pre-trained models as training in the field is possible, but cannot be relied on, and then use the iterative process to create opportunities to build and refine appropriate training sets.
Conclusion
Realising the promise of AI for Australian Defence requires a practical approach that considers the end user's context and needs. There is a temptation to implement recently published, state-of-the-art neural networks to solve every challenge. However, unless the computational needs, data access, interpretability and a myriad of other considerations are accounted for, the results may have been superseded by the time they are implemented in an Australian military context.
In our experience a team approach is likely to yield the best results, combining diverse knowledge and experience from machine learning, systems engineering, information technology, electronics engineering, industrial design and usability to ensure that we can deliver a timely and practical boost to Australian Defence capability.