Invariant Visual Object and Face Recognition: Neural and Computational Bases, and a Model, VisNet

Area of Science:

• Computational Neuroscience
• Visual Perception
• Cognitive Science

Background:

• The primate inferior temporal cortex exhibits neurophysiological evidence for invariant object and face representations.
• Understanding the computational mechanisms underlying these invariant representations is crucial for visual neuroscience.

Purpose of the Study:

• To describe a computational approach for forming invariant visual representations in the brain.
• To build a feature hierarchy model that self-organizes invariant representations based on visual input statistics.

Main Methods:

• Developed VisNet, a feature hierarchy model utilizing self-organizing learning.
• Employed associative synaptic learning with short-term memory (temporal continuity) and/or continuous spatial transformation learning (spatial continuity).
• Extended the model to account for dorsal visual system functions, attention, object selection, and scene representation.

Main Results:

• VisNet successfully builds object representations invariant to translation, view, size, and lighting.
• Model extensions explain invariant global motion perception in the dorsal stream.
• Further extensions address attentional control, object selection in complex scenes, and spatial scene representation.

Conclusions:

• The VisNet model provides a neurophysiologically-grounded computational framework for invariant visual representation.
• The approach accounts for a wide range of visual processing capabilities across ventral and dorsal streams.
• This work offers insights into how the brain achieves robust object and scene recognition.