Brain2Model Transfer: Training sensory and decision models with human neural activity as a teacher

Area of Science:

• Cognitive Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Transfer learning typically uses large, pre-trained artificial models.
• The human brain efficiently learns abstract representations with less data and computation.
• Artificial models often require extensive data and computational resources.

Purpose of the Study:

• Introduce Brain2Model Transfer Learning (B2M) to leverage human brain activity for training artificial neural networks.
• Propose two B2M strategies: Brain Contrastive Transfer and Brain Latent Transfer.
• Investigate the efficiency and effectiveness of brain-inspired transfer learning.

Main Methods:

• Developed B2M framework using human neural activity as teacher models.
• Implemented Brain Contrastive Transfer to align brain activity and network activations.
• Implemented Brain Latent Transfer using supervised regression of brain-derived features.
• Validated B2M in memory-based decision-making and scene reconstruction tasks.

Main Results:

• Student networks trained with B2M converged faster than those trained in isolation.
• B2M improved predictive accuracy in tested artificial neural networks.
• Demonstrated successful application in recurrent neural networks and variational autoencoders.

Conclusions:

• Human brain representations are valuable for training artificial learners.
• B2M offers a more efficient alternative to purely artificial training for complex decision-making.
• Paves the way for developing more efficient AI models inspired by neuroscience.