A feature-based generalizable prediction model for both perceptual and abstract reasoning.

Area of Science:

• Artificial Intelligence
• Cognitive Science
• Neuroscience

Background:

• Human intelligence excels at abstract rule inference and application.
• Current deep learning models struggle with rule expression and extrapolation.
• Existing AI tasks often lack perceptual reasoning capabilities.

Purpose of the Study:

• Develop an AI model for abstract rule detection and application.
• Enable one-shot inference and rule expression in AI.
• Address limitations of current AI in symbolic and perceptual reasoning.

Main Methods:

• Algorithmic approach using feature detection and affine transformation estimation.
• Application to a simplified Raven's Progressive Matrices task.
• Evaluation of symbolic and perceptual reasoning capabilities.

Main Results:

• Achieved near human-level performance on symbolic reasoning tasks.
• Demonstrated one-shot inference capabilities.
• Successfully handled perceptual challenges with continuous patterns.

Conclusions:

• The model shows promise for understanding abstract reasoning in humans.
• Results have implications for developing more intelligent machines.
• The approach allows for rule expression and multi-step predictions.