Macaques recognize features in synthetic images derived from ventral stream neurons

Area of Science:

• Neuroscience
• Computer Vision
• Primate Cognition

Background:

• The primate visual system's ability to recognize diverse images needs computational explanation.
• A hypothesis suggests visual cortex neurons learn and interpolate patterns from visual input.
• Neuronal patterns, termed 'prototypes,' are computationally instantiated using machine learning.

Purpose of the Study:

• To investigate if monkeys perceive similarities between neuronal prototypes and real-world objects.
• To determine if monkeys can classify prototypes based on their neural origins (inferotemporal vs. primary visual cortex).
• To validate findings with human participants and computational models.

Main Methods:

• Utilized a two-alternative forced choice task to assess monkey perception.
• Trained monkeys to identify synthetic 'monkey' images (prototypes).
• Classified prototypes derived from inferotemporal (IT) and primary visual cortex (V1) neurons, comparing them to random images.

Main Results:

• Monkeys more frequently classified IT prototypes as 'conspecifics' (other monkeys) than V1 prototypes or random images.
• Convolutional neural networks partially predicted monkey choices, suggesting shared representational principles.
• Monkeys demonstrated an ability to abstract general shape information from real-world objects.

Conclusions:

• Prototypes derived from inferotemporal cortex neurons are perceived by monkeys as similar to real-world objects.
• These findings support the idea that cortical neurons store interpretable abstractions of the visual world.
• The study bridges computational models, primate behavior, and neuroscience findings.