COVID-19 has changed the world we live in. It has had a horrific impact on the health of many people and ultimately cost many lives across the globe. That said, it has been immensely impressive to see scientists rise to meet challenges posed by COVID. More specifically, the successful development and deployment of COVID vaccines demonstrates what scientific communities can achieve.
The International Society of Information Fusion (ISIF) was founded in 1998 as the professional body that represents another scientific community. The ISIF membership is focused on developing techniques for fusing data from multiple sensors and/or over time to inform and make decisions. Solutions to problems involving Information Fusion have broad applicability, including specifically in the context of combining data to infer how the COVID epidemic will evolve but also, for example, in the context of processing data related to autonomous cars and in helping decision makers to respond to disparate, incomplete, ambiguous and conflicting information.
The environment that ISIF inhabits has changed since 1998: the minimum spec to run Windows 98 was a 66 MHz “486” PC with 16Mb of memory. Moore’s law is frequently quoted as stating that processor speed doubles every 2 years. Were that the case, the minimum spec for a 2021 release of Windows should be approximately 200 GHz. Of course, it is less: CPU processor speed has plateaued and Moore’s law is now delivered via access to many-cored processors (eg GPUs). Against this context, as a result of being driven by commercial ambition, rather than being limited by current semiconductor design, the Deep Learning community is using a number of transistors that doubles every four months (with the compute involved in training the state-of-the-art GPT-3 network estimated to cost £4.6M and to involve calculations equivalent to using 355 GPUs for a year). The Information Fusion community needs to make more extensive use of such large-scale computing resources, eg to estimate the parameters of non-linear models that can maximise performance when combining multi-source navigation data or to use off-line learning to perform on-line optimisation of the sequence of sensor management actions that maximise their utility when measured over strategic timescales.
Neighbouring communities, including but not limited to Deep Learning, make extensive use of benchmark problems with accompanying metrics to ensure that the improvements claimed in any publication are relative to the performance of the pre-existing state-of-the-art on the same problem. It is all too easy for a Information Fusion researcher to mis-implement a competitor approach or mis-configure a baseline. Doing so leads to futile debates, wastes time and delays genuine progress. There are efforts in ISIF, notably the “Evaluation of Techniques for Uncertainty Representation Working Group” (ETURWG), that are focused on producing benchmark problems and metrics. We need to adopt the specific output from this working group (eg in the context of developing solutions to pressing high-level fusion challenges) and its approach more generally (eg by helping forward-thinking defence organisations to see the benefit of releasing benchmark datasets (real and simulated as well as with both ground-truth and relevant metrics), that could then be the focus of special sessions at Fusion conferences).
Alongside benchmark problems, neighbouring communities contribute to and make extensive use of open-source frameworks. Particularly, if the modular design of the framework mirrors the modular nature of the underpinning maths, adopting open-source development enables researchers to build on the shoulders of giants and to explicitly advance the state-of-the-art. Open-source development also eases the process of practitioners making use of the outputs from previous research. It is all too easy to confuse learning about tracking by implementing algorithms from scratch with activities that could (and arguably should) be contributions to open-source frameworks (eg enhancing the state-of-the-art by developing a slight twist on a pre-existing algorithm or developing a valuable piece of IP that could be integrated into each of many algorithms). Again, there are efforts in ISIF, notably as part of the “Open-Source Tracking and Evaluation Working Group” (OSTEWG) which are seeking to develop such open-source instantiations of Data Fusion. We need to work together to contribute specifically to, in OSTEWG’s context, Stone Soup. More generally, we need to work together to design, develop and deploy such open-source articulations of the ISIF membership’s collective understanding of Data Fusion, eg such that we have something like Stone Soup for high-level fusion.
ISIF could improve by embracing modern computing resources, benchmarks and open-source software. However, any such step to the future will need to also capitalise on ISIF’s heritage. ISIF is a community that understands how to manage uncertainty, combine information across different sensors, accumulate information over time and inform decision making by both humans and machines. We need to enable that understanding to make a difference to the world we live in.
As ISIF President for 2021 and 2022, I plan to encourage the ISIF memberships’ adoption of modern computing resources, benchmarks and open-source software. I’d therefore welcome engagement with any current or potential ISIF members, who want to help ISIF move in those directions. I also plan to make sure that the community which ISIF represents is supported by the transparent operation of ISIF’s Board of Directors (BoD) and that the ISIF BoD adopts (and documents) strategies that will enable our community to achieve ISIF’s full potential.
ISIF President, 2021 and 2022.