Simple and complex cells revisited: toward a selectivity-invariance model of object recognition

Area of Science:

• Computational Neuroscience
• Computer Vision

Background:

• The ventral stream is crucial for object recognition.
• Existing models often struggle to balance selectivity and invariance.
• Revisiting the computational abstraction of simple and complex cells offers a new perspective.

Purpose of the Study:

• To propose a novel theoretical framework for modeling ventral stream processing.
• To develop a hierarchical sparse coding approach for object recognition.
• To disentangle selectivity and invariance computation in visual processing.

Main Methods:

• Organizing the perspective into three levels: computational, algorithmic, and implementation.
• Abstracting simple and complex cells into space partitioning and composition.
• Developing a hierarchical extension of sparse coding exploiting manifold constraints.

Main Results:

• A new computational abstraction for simple and complex cells based on redundancy exploitation.
• An over-parameterized hierarchical model for object recognition.
• Disentanglement of selectivity and invariance computation, differing from existing models.

Conclusions:

• The proposed framework offers a new theoretical perspective on ventral stream processing.
• The hierarchical construction can be interpreted as cortically local subspace untangling.
• Potential implementations include asymmetric sparse autoencoders and divergent spiking neural networks.