Deep Neural Networks for Modeling Visual Perceptual Learning.

Area of Science:

• Computational Neuroscience
• Machine Learning in Vision Science
• Neural Plasticity Research

Background:

• Existing models of visual perceptual learning (VPL) offer superficial connections between behavior and brain plasticity.
• Current VPL models struggle to adapt to novel stimuli and training paradigms, limiting their explanatory power.
• There is a need for more complex computational models that capture the multi-stage nature of VPL.

Purpose of the Study:

• To investigate whether a deep neural network (DNN), not specifically designed for VPL, can serve as a computational model for VPL.
• To analyze the DNN's ability to reproduce key behavioral and physiological findings in VPL research.
• To explore the DNN's potential as a test bed for VPL theories and a generator of new research predictions.

Main Methods:

• A pre-trained deep neural network (DNN) was trained on a Gabor orientation discrimination task.
• The DNN's performance was analyzed for behavioral learning characteristics, including specificity and transfer.
• Plasticity distribution within the DNN's layers was examined and compared to electrophysiological data from primate visual areas.

Main Results:

• The DNN successfully reproduced key behavioral VPL results, such as increased specificity with task precision.
• The model demonstrated asymmetric transfer from precise to coarse discriminations and showed plasticity shifts towards lower layers with increased precision.
• Learning patterns in the DNN units closely resembled electrophysiological recordings from monkey visual areas, validating tuning changes.

Conclusions:

• Deep neural networks (DNNs) provide a complex and adaptable computational model for studying visual perceptual learning (VPL).
• The DNN successfully replicated behavioral and physiological patterns, fulfilling predictions of existing VPL theories.
• This approach offers a novel method for investigating VPL across multiple analytical levels, from behavior to neural correlates.