Neural Prediction Errors as a Unified Cue for Abstract Visual Reasoning.

Area of Science:

• Artificial Intelligence
• Neuroscience
• Cognitive Science

Background:

• Deep neural networks struggle with abstract visual reasoning (AVR), hindering artificial general intelligence.
• Prediction errors are a fundamental concept in neuroscience for learning and adaptation.

Purpose of the Study:

• To propose prediction errors as a unified mechanism for supervised and self-supervised learning in AVR.
• To develop novel computational models for AVR inspired by neuroscience.

Main Methods:

• A supervised learning model framing AVR as prediction-and-matching, using prediction error (discrepancy) between predicted and candidate features.
• A self-supervised model where prediction errors unify learning and inference.

Main Results:

• Both supervised and self-supervised prediction-based models achieved state-of-the-art performance across diverse AVR datasets.
• Hierarchical prediction errors in the supervised model decreased during training, mirroring biological dopamine signal changes.

Conclusions:

• Prediction errors play a critical role in enabling abstract visual reasoning in AI.
• Leveraging neuroscience principles can advance computational models for high-level cognitive functions.