Learning the selectivity of V2 and V4 neurons using non-linear multi-layer wavelet networks

Area of Science:

• Computational Neuroscience
• Computer Vision
• Neuroscience

Background:

• Natural images contain statistical dependencies that the primate visual system exploits.
• Higher visual areas (V2-V4) are thought to process these statistical properties.
• Understanding neural information processing requires models that capture these complex interactions.

Purpose of the Study:

• To develop and analyze a non-linear network model for primate visual cortices (V2-V4).
• To investigate the emergence of selectivity and invariance properties in neural information processing.
• To explore non-linear interactions crucial for exploiting higher-order statistical redundancies in natural images.

Main Methods:

• Optimization of a two-stage non-linear network to reduce statistical dependencies in natural images.
• Analysis of network unit selectivity and invariance properties.
• Comparison of model unit selectivity with electrophysiological data from V2 neurons.

Main Results:

• The network model exhibits units with high selectivity and stimulus class invariance.
• The model's selectivity histogram shows a stronger tendency towards higher selectivities than observed in V2 neurons.
• Emergent non-linear interactions between coefficients at different scales and orientations were identified.

Conclusions:

• The developed non-linear network effectively models aspects of neural information processing in primate visual cortices.
• The model's properties, including enhanced selectivity and non-linear interactions, are crucial for processing natural image statistics.
• The findings support the role of higher-order statistical redundancies in visual information processing and suggest avenues for multi-layer system extensions.