Diverse task-driven modeling of macaque V4 reveals functional specialization towards semantic tasks

Area of Science:

• Neuroscience
• Computer Vision
• Computational Neuroscience

Background:

• Area V4 is a mid-level visual processing area.
• Convolutional Neural Networks (CNNs) trained on image classification predict V4 responses.
• The role of V4 in tasks beyond object classification is unclear.

Purpose of the Study:

• Investigate normative accounts of V4 function.
• Compare V4 and V1 responses to natural images.
• Determine the role of diverse computational objectives in V4 function.

Main Methods:

• Predicted macaque single-neuron responses in V4 and V1.
• Used representations from 23 CNNs trained on various computer vision tasks (semantic, geometric, 2D, 3D).
• Assessed model performance based on task type and selectivity.

Main Results:

• V4 responses were best predicted by semantic classification features.
• V4 showed high task selectivity, unlike V1.
• Non-semantic tasks explained aspects of V4 function not captured by individual semantic tasks.
• Combining semantic tasks improved V4 model performance.

Conclusions:

• V4's primary functional role is likely semantic processing.
• V4's selectivity to 2D/3D stimulus properties may arise from semantic goals.
• CNNs trained on diverse tasks offer insights into V4's computational objectives.