Neural and computational mechanisms underlying one-shot perceptual learning in humans

Area of Science:

• Neuroscience
• Artificial Intelligence
• Cognitive Science

Background:

• Human rapid learning and generalization are key cognitive abilities.
• One-shot perceptual learning, a rapid visual adaptation after single exposure, remains poorly understood.
• Identifying the neural basis of one-shot learning is crucial for AI and neuroscience.

Purpose of the Study:

• To pinpoint the brain regions responsible for one-shot perceptual learning.
• To elucidate the neural mechanisms, including plasticity, that support this ability.
• To develop and validate an AI model that replicates human one-shot learning.

Main Methods:

• Utilized psychophysics to measure human perceptual changes.
• Employed 7 Tesla functional Magnetic Resonance Imaging (7T fMRI) for high-resolution brain activity.
• Incorporated intracranial recordings for direct neural data.
• Developed a deep neural network (DNN) vision transformer with top-down feedback.

Main Results:

• Identified the high-level visual cortex as the primary neural substrate for one-shot perceptual learning.
• Demonstrated that neural plasticity in this region underpins the learning process.
• The developed DNN model successfully mimicked human behavioral outcomes.
• The AI model's learned prior knowledge closely resembled neural codes in the human visual cortex.

Conclusions:

• One-shot perceptual learning in humans is supported by neural plasticity in the high-level visual cortex.
• A neuro-inspired deep learning model can effectively replicate and predict human one-shot learning.
• This research provides insights into the neurocomputational mechanisms of rapid visual learning.