AI models can manipulate time to make better predictions across a wide range of areas

In the world around us, there is much in the context of time. The path of a bird through the sky is understood as a different position over a period of time, and the series of words that occur one after the other is understood.

Computer scientists and statisticians call the time series of these sequences. Statisticians have found ways to understand these patterns and predict the future, but modern deep learning AI models struggle to function just as well, if not worse than statistical models.

Engineers at the University of California, Santa Cruz have developed a new method of time series prediction that promotes deep learning that can improve predictions based on data in the near future. When applying this approach to the key task of seizure prediction using EEG data, they found that their strategy improved seizure performance by up to 44.8% to predict seizures compared to baseline methods. Although they focused on this important healthcare application, researchers' methods have been designed to be relevant to a wide range of fields. New research on Natural Communication I'll report on their results.

“The world is a chronological machine, everything is constantly changing,” says Jason Eschragian, assistant professor of electrical and computer engineering and a senior author of the study. “If you want AI to do something good for the world, you need to be able to process dynamic, ever-changing information that changes with the times.”

Future guides, explanations

Skyna Sekaran, an undergraduate researcher at Baskin Engineering who led the study, appeared in her first year at UC Santa Cruz, who is already interested in deep learning. She quickly joined Eshraghian's Neuromorphic Computing Group, which featured the project to improve deep learning time series forecasting.

The researcher's “Future-Guided Learning” method works with two deep learning models to predict time series, but with different timescales.

The two models act together as “students” and “teachers.” Teachers need to use more data as they are in a relative future and closer to events that are asked to predict. The teacher returns the results of the prediction back to the students. This knowledge transfer improves the students' ability to further predict the future.

In the context of seizure data, this method works as follows: Teachers are asked to look at the brain wave data and detect whether a seizure is occurring at that exact moment. The student model, which has been in the past 30 minutes, is trying to predict whether a seizure will occur in 30 minutes. Teacher information about whether a seizure is currently occurring is passed to students so that students can learn from the teacher's “future” perspective. Everything we learn in real time will continuously improve our students' ability to predict future seizures.

“If the seizure detection model is soaring, there's a very high chance that a seizure is occurring, if a student model is triggered that will trigger a student model that is trying to predict 30 minutes in the future. This makes the model come true because students know that they're saying, 'Hey, there'll be a seizure during this period.'

Personalized medical care

This technique provides personalized medical opportunities as students and teachers can work together to make predictions based on the unique brain signal patterns of individual patients. Combining deep learning with wearable technology is one way to enable this.

“Imagine wearing a smartwatch that can track EEG signals. Then the signal is passed to two different models. Do you have a seizure right now? Do you have a seizure in the future? You can do guided learning in that situation.

Improved seizure prediction

The researchers tested their methods with two sets of EEG data measuring brain activity from actual seizure patients. One of the datasets at Boston MIT in Children's Hospitals led researchers to improve predictive performance by 44.8%, with teacher and student models making both learning and predictions from individual patient patterns.

Data from the American Epilepsy Association trained the teacher model on generalized seizure data rather than on specific patients. Even though this more generalized version more closely represents real-world situations, this method provided an 8.9% improvement in performance over the baseline method.

The researchers also tested their methods with a common benchmarking task that engineers use to test the ability of AI models to predict complex systems known as the Mackey-Glass equations. In this mathematical application, the method showed a performance of 23.4% compared to the baseline model.

I was inspired by the brain

Eshraghian's work on deep learning takes inspiration from the miraculous ability of the human brain to process vast amounts of information with relatively little energy when AI models are expensive, painful and energy-intensive.

“We need to get inspiration from areas other than just building improvement,” Gunasekaran says. “We need to look at the brain and theory of cortical function and improve the way we use deep learning methods to make time series predictions.”

In this case, researchers draw inspiration from the way the brain acts as a predictor and always infer the next bit of sensory information it experiences. If reality is different from what you predict, you simply handle the “surprise.”

“It is this idea that perception and behavior continue to be coordinated together to minimize that error. So, in a way, when you are surprised, it helps you learn faster,” Eshraghian said.

Now, researchers may learn more about how the brain makes predictions on different timescales using results on deep learning.

“This could provide insight into how the brain uses time to calculate. The next logical step in this study will explore the dynamic space of time and see how the brain adapts to predictions across timescales,” says Eshraghian.