In the world around us, many things exist in the context of time: a bird's path through the sky is understood as different positions over a period of time, and conversations as a series of words occurring one after another.
Computer scientists and statisticians call these sequences time series. Although statisticians have found ways to understand these patterns and make predictions about the future, modern deep learning AI models struggle to perform just as well, if not worse, than statistical models.
Engineers at the University of California, Santa Cruz, developed a new method for time series forecasting, powered by deep learning, that can improve its predictions based on data from the near future. When they applied this approach to the critical task of seizure prediction using brain wave data, they found that their strategy offers up to 44.8% improved performance for predicting seizures compared to baseline methods. While they focused on this critical healthcare application, the researchers' method is designed to be relevant for a wide range of fields. A new study in Nature Communications reports their results.
"The world is a time series machine, everything is constantly changing," said Jason Eshraghian, assistant professor of electrical and computer engineering and the study's senior author. "If you want AI to do something good for the world, you need to be able to handle time-varying dynamic, ever-changing information."
Future-guided, explained
Baskin Engineering undergraduate student researcher Skye Gunasekaran, who led this research, came into her first year at UC Santa Cruz already interested in deep learning. She quickly joined Eshraghian's Neuromorphic Computing Group and took up this project to improve deep learning time series forecasting.
The researchers' "future-guided learning" technique operates with two deep learning models that work together, both making time series predictions but on different timescales.
The two models act together as a "student" and "teacher." The teacher is in the relative future, closer to the event it is being asked to predict, and therefore has more data to work with. The teacher passes the results of its predictions back to the student—whether it was correct or made an error. This knowledge transfer improves the ability of the student to make its predictions further into the future.
In the context of seizure data, the method works as follows: the teacher is asked to look at brain wave data and detect whether a seizure is occurring at that exact moment. The student model, which is 30 minutes in the past, is attempting to predict if a seizure will occur in 30 minutes. The teacher's information about whether the seizure is happening currently is passed back to the student, so that the student can learn from the teacher's 'future' perspective. Everything learned in real-time continuously improves the student's ability to predict future seizures.
"If the seizure detection model spikes, saying there's a very high probability that a seizure is occurring, that triggers the student model, which is trying to predict 30 minutes into the future. That student now knows that the teacher is saying, 'Hey, there's going to be a seizure in this time period,' so it triggers the model, achieving better accuracy. It learns to associate its current input data with the seizure," Gunasekaran explained.
Personalized medicine
This technique provides an opportunity for personalized medicine, because the student and teacher can work together to make predictions based on the unique brain signal patterns of an individual patient—when asking a doctor to provide constant, individualized feedback is very costly. Combining deep learning with wearable technology could be one way to enable this.
"Imagine you're wearing a smart watch which is able to track your EEG signal. Then, that signal gets passed to two different models: is there a seizure right now, and is there going to be a seizure in the future? You could perform future guided learning in that situation, where it would be actively beneficial—the model can learn with a lot of time and a lot of data," Gunasekaran said.
Better seizure predictions
The researchers tested their method with two sets of EEG data, which measures brain activity, from real seizure patients. With one of the datasets, from the Children's Hospital Boston MIT, the researchers saw a 44.8% improvement in prediction performance, with the teacher and student model both learning and making predictions from the patterns of an individual patient.
With data from the American Epilepsy Society, the teacher model was trained on generalized seizure data rather than that of the specific patient. Even with this more generalized version, which more closely represents a real-world situation, the method saw a 8.9% improvement in performance over baseline methods.
The researchers also tested their method with a common benchmarking task that engineers use to test their AI models' ability to make predictions about complex systems, called the Mackey-Glass equation. In this mathematical application, their method showed 23.4% better performance compared to baseline models.
Brain-inspired
Eshraghian's research on deep learning draws inspiration from the human brain's miraculous ability to process huge amounts of information with relatively little energy—in a time when AI models are expensive and painfully energy intensive.
"We need to get inspiration from areas other than just architectural improvements," Gunasekaran said. "We need to possibly look at the brain and theories of cortical function to try and improve the way that we can use deep learning methods to do time series forecasting."
In this case, the researchers draw inspiration from the way that the brain acts as a predictive machine, constantly guessing what the next bit of sensory information it will experience. It only processes "surprises"—when reality is different from what it has predicted.
"It's this idea that perception and action keep adjusting together to minimize that error. So, in a way, being surprised will help you learn faster," Eshraghian said.
Now, the researchers might use their results about deep learning to learn more about how the brain makes predictions on different timescales—from milliseconds to months ahead.
"This could give us more insight about how the brain uses time to compute—the next logical steps for this research could be exploring the dynamic space of time, and seeing how the brain adapts to predictions across timescales," Eshraghian said.
