A connectionist architecture for view-independent grip-aperture computation

Area of Science:

• Computational Neuroscience
• Computer Vision
• Cognitive Science

Background:

• Recognizing object-directed actions requires robust visual features invariant to viewpoint.
• Grip aperture is a key feature for understanding reach-to-grasp actions.
• Existing models lack view-invariant feature extraction for action recognition.

Purpose of the Study:

• Introduce a computational model (NeGOI) for extracting view-invariant grip aperture.
• Investigate the neural processing of visual features for action recognition.
• Provide biologically plausible, high-level visual features for mirror systems.

Main Methods:

• Developed NeGOI (neural network architecture for measuring grip aperture in an observer-independent way).
• Utilized view-independent units (VIP units) selective for prototypical hand shapes.
• Integrated VIP unit outputs to compute grip aperture from superimposed hand shapes.

Main Results:

• NeGOI successfully extracts grip aperture in a view-independent manner.
• The model demonstrates effective grip aperture recognition properties.
• NeGOI's architecture aligns with functional models of the ventral visual stream (up to STS).

Conclusions:

• NeGOI provides a parsimonious set of high-level, view-independent visual features.
• The model offers a biologically plausible architecture for action recognition.
• NeGOI's features serve as input for mirror systems, enhancing understanding of action processing.