AI Models Predict Biological Learning Patterns in Mouse Brains
University of Utah researchers find that artificial neural networks and living brains share identical learning trajectories and failure modes.
Artificial intelligence and biological brains learn—and fail—in strikingly similar ways, according to new research from the University of Utah. By comparing machine learning models with living mice, scientists discovered that both systems require a specific sequence of experience to master complex tasks, mirroring each other in both performance and internal neural activity.
The study, published in Nature Neuroscience, found that both neural networks and living brains show significantly improved performance on complex tasks when they are first trained on a simpler version of that task. Conversely, when researchers skipped these foundational steps and jumped straight to complex requirements, both the AI and the animals made the same predictable errors, such as responding too early to "go" trials. Beyond behavioral outcomes, the researchers observed that large-scale neuron firing patterns in the mice closely mirrored the activity patterns adopted by the trained neural networks.
The Role of the Medial Entorhinal Cortex
To reach these conclusions, the research team focused on the medial entorhinal cortex, a region of the brain critical for learning. The experimental design utilized a dual-model approach: a biological model involving mice rewarded for responding to timed patterns of smells, and a virtual neural network trained on a corresponding "go" versus "no-go" stimulus task. This parallel structure allowed researchers to isolate how structured experience shapes the way a system develops a strategy for solving a problem.
Implications for Neuroscience and AI
This alignment suggests that computational models can serve as accurate predictors for biological learning behaviors and neural dynamics. By modeling the process up front, the team was able to move more quickly toward useful hypotheses while reducing the number of animals required for testing. At an abstract level, the AI serves as a strong analogy for how the brain is thought to function.
Future Applications and Cognitive Health
The ability to simulate biological learning in a virtual environment opens new doors for both education and medicine. By understanding the fundamental principles of how the brain acquires skills, researchers believe they can better identify why these processes fail. This understanding is vital for addressing cognitive failures in diseases such as Alzheimer’s, providing a roadmap for how to potentially fix broken neural dynamics.
Moving forward, scientists will look to expand these models to see if similar patterns hold across different brain regions or more diverse species. While the mirroring between the medial entorhinal cortex and artificial networks is clear, the extent to which this applies to higher-order human cognition remains a primary area for future investigation.