Atoms of recognition in human and computer vision

Area of Science:

• Neuroscience
• Computer Vision
• Cognitive Psychology

Background:

• Object recognition is a fundamental challenge in visual neuroscience.
• Neural network models show progress in visual object recognition, mimicking human performance.
• Current models' representations and learning processes may not fully align with human visual systems.

Purpose of the Study:

• To investigate if current neural network models use similar representations and learning processes as the human visual system.
• To identify critical visual features and processes used by humans but not by current models.
• To enhance computational models for improved visual recognition and interpretation.

Main Methods:

• Introduction and use of minimal recognizable images in psychophysical studies.
• Analysis of human sensitivity to precise feature configurations in minimal images.
• Simulations using current neural network models to assess their recognition of minimal images.

Main Results:

• Human visual system utilizes specific features and processes not employed by current AI models.
• Minute changes in minimal recognizable images drastically affect human recognition, highlighting critical features.
• Current models fail to explain this sensitivity and do not achieve human-level recognition of minimal images.

Conclusions:

• Human visual recognition relies on features uniquely revealed at the minimal recognizable image level.
• Current computational models lack the capacity to learn and utilize these essential features.
• Understanding these features is crucial for advancing visual recognition theories and AI capabilities.