UK scientists claim AI efficiency breakthrough for geospatial analytics

New model can crunch massive datasets without the high power demand

Scientists from the University of Glasgow claim a major advance in geospatial modelling, which they say significantly reduces power demand while retaining accuracy.

As helpful as it can be in crunching through large datasets, in real-world use cases machine learning often hits bottlenecks that make modelling scenarios extremely computationally (and therefore financially and environmentally) expensive.

Examples include the processing of geospatial data to forecast local air pollution levels, identify house price trends or analyse pockets of poverty, where the size and diversity of the data can make it hard to produce consistently accurate results.

Recently, advances have been with neural networks, which can more accurately capture complex spatial relationships at scale than older techniques such as spatial and geo-statistical modelling.

Transformer models operating over a structure that represents datapoints as a graph improve accuracy by aggregating information from neighbouring points on the graph according to relation-based attention scores, making them more sensitive to positional information.

But, say researchers from Glasgow University and Florida State University, the generalisation and practical application of these models is hindered by the fact that "computational and time complexity increasing quadratically with input sequence length." In other words, the compute power required increases exponentially with the size and variability of the dataset. Unfortunately making sense of large heterogenous datasets is part and parcel of geospatial analyst’s job.

To tackle this energy use problem the researchers have built a new transformer-based model called GeoAggregator, enabling accurate processing of even very large and varied geospatial datasets without the associated hunger for power. It analyses spatial autocorrelation (how nearby places influence each other) and spatial heterogeneity (how patterns vary from one location to another) with far lower demands on resources than existing models.

The team claim two main innovations for this model. First, a "Gaussian-biased local attention mechanism" helps the model selectively focus on relevant nearby data points based on local proximity. Second, a "Multi-head Cartesian Product Attention (MCPA) mechanism", which they say keeps the model lightweight while maintaining high accuracy.

In tests, the model proved equal to or better than competitive models in terms of accuracy across a range of tasks and datasets, while using significantly fewer computational resources — some three times less than equivalent models without MCPA and orders of magnitude less than some against spatial statistical models and state-of-the-art geospatial deep learning methods.

Results are published in a preprint paper which will be presented at the AAAI Conference on Artificial Intelligence this week. The code has been open-sourced and published on GitHub.

Lead author Rui Deng said: “For small and medium-sized companies, researchers or teaching purposes where resources are limited, Aggregator provides a way to get highly accurate data analysis while maintaining efficiency. Even for larger organisations with unlimited computational resources, choosing a more efficient model like this one could boost their efforts to achieve sustainability through reduced energy and water consumption."