Tipping points are all around us, but it’s hard to see them coming. When a person falls off a bike, a patient has a heart attack, or a campfire escalates to a wildfire, it is nearly impossible to identify the exact moment at which disaster became inevitable.
But it might not be impossible for much longer.
On Sept. 28, Thomas Bury, a researcher in McGill’s Department of Physiology, and Chris Bauch, a professor at the University of Waterloo, published a paper detailing the development of a new deep learning (DL) algorithm, that identifies early warning signals (EWS) of tipping points in dynamical systems.
Dynamical systems are all around us and are defined as systems in which many variables interact and evolve over time. From the Earth’s climate to the human body, the organizations and organisms that dictate our quality of life are complex and unpredictable.
DL algorithms can improve their output automatically through experience and the use of data. But even with the help of DL systems, there remains the problem of not having enough data. Artificial intelligence algorithms need to be trained on data sets in order to learn how to recognize and predict patterns. Many of the areas researchers hope to explore, such as climatic dynamics or heartbeats, do not yet have enough data to properly train an algorithm.
However, data does not need to be taken from field experiments for it to be an accurate reflection of real events. Previous research into dynamical systems revealed key patterns about how data changes close to tipping points. By simplifying these patterns and building computer models to represent them, it is possible to generate vast amounts of simulated data that reflect real systems without actually collecting observations in the field.
Using simulated data representing dynamical systems with and without tipping points, Bury and Bauch trained their DL algorithm to recognize which features corresponded to systems with tipping points. Once trained, the algorithm was able to examine new situations, real and simulated, and successfully make predictions about the likelihood of future tipping points.
Bury, who has been studying tipping points since 2015, is interested in early warning signals in both ecological and medical contexts.
“I am inspired to study tipping points because they arise in such diverse areas of science and society,” Bury said in an interview with The McGill Tribune. “This algorithm, for the first time, combines deep learning and dynamical systems theory to predict tipping points, and does so with better accuracy than previous approaches.”
Although their algorithm is still undergoing testing to avoid unexplained inaccuracies, the possible applications of such an algorithm are endless. The ability to predict regional tipping points in climate systems could enable governments to prepare for natural disasters, such as floods or droughts, and better adapt to the effects of climate change.
The algorithm can also be applied to unpredictable events at the individual level. For patients with cardiac arrhythmia, a condition that heightens the risk of sudden death by heart attack, data taken from heart monitors could be used to predict if and when a patient might experience a heart attack.
“Our approach may provide individuals [and] societies with greater forewarning to these events, and therefore allow mitigative [and] preventative strategies to be implemented in advance of the tipping point,” Bury said.
In the world of mathematics, Bury and Bauch are some of the first to study deep learning and dynamical systems simultaneously.
“I find this research particularly exciting as it has shown that the combination of two seemingly disparate areas of mathematics, deep learning and dynamical systems, provides better prediction of tipping points than either area […] has managed on its own,” Bury said.
Though researchers cannot yet predict the next drought or medical emergency with absolute accuracy, it seems that it’s now only a matter of time before it can.