Straddling neuroscience and computer science, an EPFL researcher has designed an artificial intelligence (AI) algorithm capable of predicting the effects of surprise or novelty on behavior. A tool that could be useful in psychiatry or in education, for example.
IMAGO/Design Pics
In his doctoral thesis, Alireza Modirshanechi, researcher at the Computational Neuroscience Laboratory at the Swiss Federal Institute of Technology in Lausanne (EPFL), designed an algorithm described as an intelligent artificial agent that imitates humans. Subjected to the same experiments, he performs the same tasks with the same results.
Purpose of the operation: better define the surprise effect and study its impact on different brain functions. Using classic experiments in behavioral studies, the specialist has developed a taxonomy of 18 different mathematical definitions of surprise and novelty.
Mr. Modirshanechi then explored the similarities of these definitions, their differences and the conditions that make them indistinguishable. Its algorithm thus distinguishes surprise, considered as a modulator of the speed of learning, and novelty, a driver of exploration towards a goal.
Test predictions
“We have mathematically quantified this,” explains the researcher, quoted Friday in an EPFL press release: “We can thus distinguish that surprise accelerates the learning process, while novelty pushes exploration. We can dissociate the signals in the brain.
A second step consisted of testing the algorithm’s predictions on humans to see if they were consistent. The scientist analyzed the behavior and electroencephalogram (EEG) data of human subjects in cognitive experiments. “We were able to predict between 60 and 80% of the decisions that the subjects were going to make during the experiments,” indicates the researcher.
“Everyone knows that when you drop an apple, it falls. But Newton found the formula that explains it. That’s kind of our goal. We were able to define the algorithm that predicts when and to what degree the subject is surprised and we can explain by what equation, humans learn faster when they are surprised,” he explains.
A basis for research
This algorithm constitutes a basis for further research. “For example, EEG suggests that people with schizophrenia have a different perspective on surprise than those in control groups. But we don’t know how different their perspective is,” notes the specialist.
In other fields, such as education, this base could make it possible to explore avenues for using surprise in order to strengthen the learning process or memorization.
The other contribution of this work is AI. “Most existing algorithms are based on a stable environment. We must therefore integrate these surprise signals to update our models and design more reliable and safer AI,” concludes the postdoctoral researcher.
