It is not black and white to precisely point out how neural activity changes with learning. Recently, some people think that learning in the brain or biological learning can be considered from the perspective of optimization, which is the way of learning in artificial networks such as computers or robots. A new perspective article co-authored by researchers at Carnegie Mellon University and University of Pittsburgh links machine learning with biological learning and shows that the two methods are not interchangeable, but can be used Provide valuable insights into how the brain works.
"How do we quantify the changes we see in the brain and the changes in subjects' behavior during learning," said Byron Yu, professor of electrical and computer engineering at Biomedical Engineering . "It turns out that in machine learning and artificial intelligence, there is a perfect framework in which you can learn something, called optimization. We and others in the field have been thinking about how the brain compares with this framework, which is development Used to train artificial intelligences for learning."
The optimization view indicates that during the learning process, the brain should change in a mathematically prescribed way, which is similar to when artificial neurons are trained to drive robots or play chess. How its activities change in a specific way.
“One of the things we are interested in understanding is how the learning process unfolds over time, not just looking at snapshots before and after learning happened,” explains Jay Hennig, who recently received his PhD. Graduated from Carnegie Mellon University with a major in neural computing and machine learning. "In this opinion article, we provide three main points, which are important for people in the context of thinking about why neural activity may change throughout the learning process, and these points cannot be easily explained by optimization. ”
The main points include the inflexibility of neural variability throughout the learning process,Even in simple tasks, multiple learning processes are used, and there are large-scale activity changes that are not related to the task.
University of Pittsburgh bioengineering Professor Aaron Batista said: “It’s tempting to draw inspiration from successful examples of artificial learning agents and assume that the brain must do anything. "However, a specific difference between artificial learning systems and biological learning systems is that artificial systems usually only do one thing, and they do it very well. The activities in the brain are completely different, and many processes happen at the same time. We and others have already I observed that something happened in the brain that the machine learning model could not explain."
Carnegie Mellon University and Institute of Neuroscience Biomedical Engineering Professor Steve Chase added: "We see a theme of architecture and the future. By focusing on these areas of neuroscience that can provide information for machine learning, and vice versa, our goal is to link them to an optimized view, and finally understand how learning is unfolded in the brain at a deeper level."
This work was co-authored with Dr. Emily Oby and Darby Losey, a bioengineering research faculty member at the University of Pittsburgh. CMU Neural computing and machine learning students. The team’s work is ongoing and completed in collaboration with the Center for Cognitive Neuroscience, which is an inter-university research and education program between Carnegie Mellon University and the University of Pittsburgh, using the strengths of each institution to study cognitive and neurological Mechanisms of biological intelligence and behavior.
.
