Abstract visual reasoning based on algebraic methods.

Area of Science:

• Artificial Intelligence
• Cognitive Science
• Computer Vision

Background:

• Human cognition relies on abstract pattern extraction from complex data.
• Abstract visual reasoning is crucial for machine intelligence.
• Existing neuro-symbolic methods have limitations in exploring abstract patterns and sequential sensitivity.

Purpose of the Study:

• To develop a novel relation model for end-to-end abstract visual reasoning.
• To improve the extraction of multi-granular rule embeddings and abstract patterns.
• To enhance machine intelligence's ability to understand composite images and visual sequences.

Main Methods:

• Constructed an object-centric relation model with inductive biases.
• Employed a gating fusion module for integrating object and relationship representations.
• Utilized a relational bottleneck method to separate perceptual information and abstract representations, promoting relational comparisons.
• Bridged algebraic operations and machine reasoning via the relational bottleneck to identify invariant sequences.

Main Results:

• Achieved a total accuracy of 96.8% on the I-RAVEN dataset.
• Significantly surpassed state-of-the-art baseline methods.
• Exceeded human performance, which was recorded at 84.4% accuracy.

Conclusions:

• The proposed relation model effectively extracts high-order abstract patterns from complex data.
• The relational bottleneck method is key to inducing abstract pattern extraction and bridging algebraic operations with machine reasoning.
• The model demonstrates superior performance in abstract visual reasoning tasks compared to existing approaches and human capabilities.