A Novel Methodology for Measuring the Abstraction Capabilities of Image Recognition Algorithms

Area of Science:

• Artificial Intelligence
• Computer Vision
• Machine Learning

Background:

• The proliferation of diverse artificial intelligence (AI) systems necessitates standardized methods for intelligence measurement.
• Evaluating AI intelligence provides crucial feedback for system development and refinement.
• Current methods often struggle to differentiate genuine learning from embedded solutions.

Purpose of the Study:

• To propose a novel approach for assessing AI intelligence by examining the learning process.
• To develop a method capable of distinguishing acquired knowledge from pre-written solutions in AI systems.
• To evaluate the abstraction capabilities of image recognition software.

Main Methods:

• The study examines the learning process within AI systems, focusing on image recognition software.
• A key method involves applying various transformations to objects and assessing the software's resilience.
• The approach detects if the AI system learns to adapt to transformations through repeated exposure.

Main Results:

• The proposed method was successfully tested on a basic neural network.
• The tested neural network demonstrated an inability to learn most of the applied transformations.
• The findings indicate limitations in the network's adaptive learning capabilities.

Conclusions:

• The developed method offers a viable way to test the abstraction and learning capabilities of AI.
• The approach can be applied to various image recognition systems to gauge their intelligence.
• This work contributes to the ongoing effort to create more sophisticated and adaptable artificial intelligence.