The impact on midlevel vision of statistically optimal divisive normalization in V1

Area of Science:

• Neuroscience
• Computational Neuroscience
• Computer Vision

Background:

• The primate visual cortex, specifically areas V1 and V2, exemplifies hierarchical processing.
• Understanding V2's functional properties and its interactions with V1 is crucial.
• V1's nonlinear response properties significantly shape the inputs V2 receives.

Purpose of the Study:

• To investigate the impact of different divisive normalization models in V1 on V2's feature selectivity.
• To explore how V2 pools V1 responses through unsupervised learning.
• To compare the performance of two-stage visual models on perceptual tasks.

Main Methods:

• Simulating V1 responses using canonical and statistically optimal surround normalization models.
• Applying principal component analysis to model V2's unsupervised learning of V1 responses.
• Evaluating two-stage models on object recognition and figure/ground judgment tasks.

Main Results:

• V1 population response statistics varied significantly across different normalization models.
• V2-like feature selectivity emerged with optimal and canonical V1 normalization, but not without it.
• Models with V1 surround normalization outperformed others in object recognition.

Conclusions:

• Statistically optimal V1 normalization provides advantages in midlevel vision tasks like figure/ground judgment.
• The choice of V1 normalization significantly influences V2's computational capabilities.
• Future research on midlevel visual areas should consider stimuli that leverage V1's optimal computations.