Summary: Researchers have developed an AI-inspired neural model that mirrors how the prefrontal cortex uses gating to regulate information flow between neural regions. This framework improves adaptive learning and could guide more brain-like artificial intelligence systems.
Source: Salk Institute
Teaching computers to think and learn like humans remains a central goal in artificial intelligence. Human brains excel at applying previously learned knowledge to new situations and continually refining those skills—an adaptability that machines have struggled to match. Researchers at the Salk Institute have now created a computational model that more accurately reproduces a key brain mechanism behind that flexibility.
The team built a model that captures how the prefrontal cortex uses “gating” to control the flow of information across different neural regions. By simulating network-wide gating, the model reproduces how the brain delegates and integrates information among subregions, offering insights into both biological cognition and the design of adaptive machine learning systems.
“If this model can be scaled and incorporated into more complex AI systems, it may enable faster learning and better problem-solving,” says Terrence Sejnowski, head of Salk’s Computational Neurobiology Laboratory and senior author on the study, published in Proceedings of the National Academy of Sciences on November 24, 2020.
Human and mammalian brains rapidly interpret sensory inputs—like sights and sounds—and integrate that information with prior knowledge. This lifelong capacity for transfer and continual learning is a benchmark for artificial systems. Traditionally, machine learning models trained on one task struggle to generalize that learning to related tasks without retraining each individually. Achieving broad transfer and memory retention has therefore been a major challenge.
To address this, Sejnowski’s team developed a new computational framework that mimics prefrontal cortex activity during the Wisconsin Card Sorting Test, a classical cognitive task used to assess flexibility and rule switching. In the test, participants sort cards by color, shape, or number and must adapt as the sorting rule changes. The task is not only a clinical tool for diagnosing cognitive impairment but also a common benchmark for evaluating how well models reproduce human-like cognition.
Prior models of the prefrontal cortex struggled with this test. The new framework, however, explicitly models hierarchical gating across the entire prefrontal network, allowing distinct information to be routed to specialized subregions. While gating has been studied at small scales within local cell clusters, integrating gating across a whole network yields more realistic, flexible behavior.
The resulting network performed at human levels on the Wisconsin Card Sorting Test and reproduced characteristic error patterns observed in patients with prefrontal damage. When portions of the model were lesioned, the system produced errors similar to those seen after trauma or dementia affecting the prefrontal cortex—an outcome that strengthens the biological relevance of the architecture.
“One of the most exciting outcomes is that this modeling approach clarifies how the prefrontal cortex may be organized,” says Ben Tsuda, a Salk graduate student and first author. “That has practical implications for machine learning and for understanding diseases that impair prefrontal function.”
A clearer map of how prefrontal regions coordinate could also inform clinical interventions for brain injury. For example, identifying key nodes in the gating network might suggest targets for therapies such as deep brain stimulation or other neuromodulation approaches.
“The brain still outperforms state-of-the-art deep learning in versatility and generalization across tasks governed by different rules,” says coauthor Kay Tye, professor in Salk’s Systems Neurobiology Laboratory and the Wylie Vale Chair. “This work demonstrates how information gating can drive a more accurate and capable model of the prefrontal cortex.”
Next, the researchers plan to scale the architecture to tackle more complex tasks beyond card sorting and to test whether network-wide gating consistently improves working memory and transfer in varied learning scenarios. If those efforts succeed, the approach could inspire AI systems that adapt more readily to new problems and changing environments.
Hava Siegelmann of the University of Massachusetts Amherst is listed as an additional author on the study.
Funding: This work was supported by grants from the Kavli Institute for Brain and Mind at UC San Diego, the Office of Naval Research (N000141612829), the National Science Foundation (1735004) and DARPA (W911NF1820).
About this artificial intelligence research news
Source: Salk Institute
Contact: Press Office – Salk Institute
Image: The image is in the public domain
Original Research: Closed access.
“A modeling framework for adaptive lifelong learning with transfer and savings through gating in the prefrontal cortex” by Terrence Sejnowski, Kay Tye and Ben Tsuda. PNAS
Abstract
A modeling framework for adaptive lifelong learning with transfer and savings through gating in the prefrontal cortex
The prefrontal cortex encodes and stores many distinct schemas and flexibly switches among them. Although advances in artificial neural networks trained by reinforcement learning have begun to model core processes of schema encoding and storage, it remains unclear how the brain creates new schemas while retaining and using older ones. Here, we present a neural network framework that incorporates hierarchical gating to model the prefrontal cortex’s capacity to encode and flexibly use multiple, disparate schemas. We demonstrate that gating naturally facilitates transfer learning and robust memory savings, and that lesions in the model reproduce neuropsychological impairments observed in patients with prefrontal damage. Our architecture, named DynaMoE, offers a principled framework for how the prefrontal cortex might manage the many schemas required for navigating the real world.