Object Detection, Recognition, Deep Learning, and the Universal Law of Generalization.

Area of Science:

• Cognitive Science
• Computational Neuroscience
• Artificial Intelligence

Background:

• Object detection and recognition are crucial for species survival, involving generalization across variable appearances.
• The universal law of generalization (ULoG) proposes consistent principles governing generalization across species and tasks.
• Understanding the internal representations driving generalization in artificial systems like neural networks is challenging.

Purpose of the Study:

• To test if internal representations for generalization reflect environmental properties, as suggested by ULoG.
• To investigate how deep neural networks handle natural camouflage in object recognition.
• To extend ULoG to realistic learning scenarios and analyze internal representations.

Main Methods:

• Trained a deep neural network on natural images of clear and camouflaged animals.
• Examined internal representations to understand generalization patterns.
• Extended ULoG to a realistic learning regime and developed methods to determine category prototypes.

Main Results:

• Generalization functions were found to be monotonically decreasing with appropriate category prototypes, mirroring biological systems.
• Camouflaged inputs were systematically located at the tail of these functions, not randomly distributed.
• Results support the hypothesis that ecological properties shape internal representations for generalization.

Conclusions:

• Internal representations for object recognition generalization are primarily influenced by the environment, regardless of the system's specifics.
• The extended ULoG offers a method to analyze internal representation organization and decision-making in learning systems.
• This research bridges understanding between biological and artificial object recognition systems.