How a New AI Model Mimics the Brain’s Prefrontal Cortex

Source: mohamed_hassan/pixabay

Is there merit to the familiar adage that goes, “once you learn how to ride a bike, you never forget?” Unfortunately, this is not always the case, as people suffering from neurodegenerative diseases and disorders such as dementia, Alzheimer’s disease, or traumatic brain injuries may experience memory loss, and impairments in the ability to reason, focus, and communicate. In a recent neuroscience study, researchers from the Salk Institute and the University of Massachusetts Amherst (UMass Amherst) create an artificial intelligence (AI) machine learning model to provide new insights on how the brain’s prefrontal cortex operates when it comes to adaptive lifelong learning.

Salk Institute researchers Terrence J. Sejnowski (senior author and head of Salk’s Computational Neurobiology Laboratory), Ben Tsuda, and Kay M. Tye, along with Hava T. Sieglemann at UMass Amherst, used artificial neural networks to study the brain and published their insights in the Proceedings of the National Academy of Sciences of the United States of America (PNAS). They created a model after the human brain’s prefrontal cortex to better understand its ability to flexibly encode and use multiple disparate schemas.

How and why is the brain capable of adaptive lifelong learning? The human brain has the remarkable learning ability to create new mental representations while maintaining and using prior schemas. However, the precise neural mechanisms on how the brain achieves this has not been entirely clear.

For the study, the team used the Wisconsin Card Sort Test (WCST), a psychological test used by clinical psychologists, neuropsychologists, and clinicians to identify frontal lobe dysfunction in order to help diagnose psychiatric disorders, dementia, schizophrenia, neurodegenerative diseases, mental illness, and other conditions. The test requires the participant to sequentially sort cards according to shape, color, or number, and infer the sorting rule by trial and error. It requires the prefrontal cortex to recognize rule scenarios and adapt via reinforcement signals with changing rules.

“Our architecture, which we call “DynaMoE,” provides a fundamental framework for how the prefrontal cortex may handle the abundance of schemas necessary to navigate the real world,” wrote the researchers.

DynaMoE has a gating network comprised of a singular LSTM (Long short-term memory) network. LSTM is a type of recurrent neural network (RNN) that is often used for purposes such as handwriting recognition, speech recognition, and financial services, as well as other tasks with time-series data that require classification or prediction. Unlike feedforward AI networks, recurrent neural networks use data from both the current input as well as from recent past decisions that act like “memories” in order to formulate new responses.

Recurrent neural networks identify long-term dependencies which are the correlations between events separated by moments of time. In effect, these types of networks are a way to distribute weights over time. However, they tend to have short-term memories and a vanishing gradient problem where the gradients (values used to update the network weights) decrease during backpropagation through time and diminish the learning. The earlier neural network layers experience smaller gradient updates, which results in the recurrent neural network “forgetting” what it has seen in longer sequences.

To address this memory issue, LSTMs are a type of recurrent neural network that have gates that can learn which data in a sequence should be “remembered” to help control the flow of information.

The team also demonstrated how neuropsychological dysfunction seen in patients with brain damage in the prefrontal region are simulated using the AI model. The researchers observed that lesions to targeted functional components of the DynaMoE network resulted in various error modes in the Wisconsin Card Sort Test, similar to the error modes of patients with various types of prefrontal damage.

“Here we propose a simple neural network framework that incorporates hierarchical gating to model the prefrontal cortex’s ability to flexibly encode and use multiple disparate schemas,” the researchers wrote. “We show how gating naturally leads to transfer learning and robust memory savings. We then show how neuropsychological impairments observed in patients with prefrontal damage are mimicked by lesions of our network.”

The biological brain has served as inspiration for artificial intelligence machine learning design, such as artificial neural networks used for deep learning. Now AI is being deployed as a tool to help unravel how the brain works. This study has opened the door to a better understanding of the prefrontal cortex, an important area of the brain. Next, the researchers plan to see whether network-wide gating gives the artificial prefrontal cortex a better working memory in all types of learning scenarios­­­. The answer could help lead to more adaptable, smarter AI one day in the future.

Copyright © 2021 Cami Rosso All rights reserved.