Learning Slowness in a Sparse Model of Invariant Feature Detection

Area of Science:

• Computational Neuroscience
• Machine Learning
• Computer Vision

Background:

• Primary visual cortical complex cells function as invariant feature detectors.
• Temporal slowness and sparsity are hypothesized to be crucial for developing invariant feature detectors.

Purpose of the Study:

• To propose a unified model for learning complex cell connectivity.
• To integrate temporal slowness and sparsity for invariant feature detection.

Main Methods:

• Developed the generative adaptive subspace self-organizing map (GASSOM), extending the adaptive subspace self-organizing map (ASSOM).
• GASSOM incorporates a generative input model and learns temporal slowness as an emergent property.
• Applied GASSOM to unlabeled naturalistic image sequences from a realistic eye movement model.

Main Results:

• Temporal slowness emerged naturally from the GASSOM structure, not as an imposed criterion.
• GASSOM does not require explicit segmentation of input data into episodes.
• The model demonstrated improved invariance of feature detectors trained on image sequences.

Conclusions:

• The GASSOM provides a novel framework for learning invariant feature detectors.
• Emergent temporal slowness is key to achieving invariance in visual processing models.
• This approach advances understanding of complex cell function and visual system development.