Variability in training unlocks generalization in visual perceptual learning through invariant representations.

Area of Science:

• Cognitive Neuroscience
• Computational Neuroscience
• Visual Perception

Background:

• Visual perceptual learning is typically stimulus and location specific, limiting its practical applications.
• Generalization is crucial for applying learned skills in real-world scenarios.
• Existing learning paradigms often fail to achieve robust generalization in visual tasks.

Purpose of the Study:

• To investigate if introducing stimulus variability can promote generalization in visual perceptual learning.
• To explore the underlying mechanisms of variability-induced generalization using deep neural networks.
• To identify potential modifications for training paradigms to enhance practical applications of perceptual learning.

Main Methods:

• Trained human subjects in orientation discrimination with varying levels of spatial frequency variability.
• Assessed generalization of learning to novel stimuli and locations.
• Utilized deep neural networks trained with variable inputs to model the learning process.

Main Results:

• Stimulus variability enabled generalization to new stimuli and locations, irrespective of task difficulty.
• The amount of learning on the primary task was not negatively affected by variability.
• Deep neural networks trained with variable inputs developed invariance to task-irrelevant features, predicting generalization.

Conclusions:

• Introducing variability in task-irrelevant features is an effective strategy to achieve generalization in visual perceptual learning.
• Learned invariance in neural representations explains the mechanism behind variability-induced generalization.
• This finding suggests new avenues for understanding higher-order visual cortex function and optimizing training paradigms for practical use.