Inference via sparse coding in a hierarchical vision model.

Area of Science:

• Computational neuroscience
• Computer vision
• Visual cortex modeling

Background:

• Sparse coding offers computational benefits and biological plausibility for visual cortex models.
• The precise impact of sparsity levels on visual task performance remains unclear.

Purpose of the Study:

• To investigate how controlled sparsity levels in sparse coding affect visual tasks within a hierarchical V2 model.
• To compare sparse coding with independent component analysis (ICA) in a V2 model context.

Main Methods:

• Integrated explicit sparse coding with controllable sparsity into an existing hierarchical V2 model, replacing ICA.
• Evaluated model performance on mid-level vision tasks: figure-ground classification, texture classification, and angle prediction.
• Assessed texture sensitivity and deleted-region inference capabilities, comparing with V2 data.

Main Results:

• Higher sparsity in sparse coding basis functions yielded qualitatively different structures (curves, corners).
• Sparse coding models showed lower image classification accuracy than ICA models.
• Sparse coding models, particularly with increased sparsity, better matched V2 texture sensitivity and performed superiorly on deleted-region inference tasks.
• Increased sparsity enabled inference over larger deleted image regions.

Conclusions:

• Controlled sparsity in visual cortex models is crucial for matching biological texture sensitivity and enabling robust deleted-region inference.
• While potentially less optimal for raw image classification, sparse coding's inferential capabilities enhance mid-level vision understanding.
• The study elucidates the mechanism behind sparse coding's inference capability in visual processing.