Competition improves robustness against loss of information

Area of Science:

• Computational neuroscience
• Artificial intelligence
• Computer vision

Background:

• Modeling primary visual cortex (V1) receptive fields is crucial for understanding early vision.
• Existing models often focus on similarity to biological data, but robustness against information loss is underexplored.

Purpose of the Study:

• To investigate the influence of competition mechanisms on the robustness of early vision models against occlusions.
• To compare different competition mechanisms for their effectiveness in enhancing model resilience.

Main Methods:

• Four methods with distinct competition mechanisms were compared: independent component analysis, non-negative matrix factorization, predictive coding/biased competition, and a Hebbian network.
• Model robustness was evaluated using classification accuracy on the MNIST handwritten digit dataset with varying levels of occlusion.

Main Results:

• Methods employing competitive mechanisms demonstrated higher robustness against information loss due to occlusions.
• Predictive coding with global feedback inhibition showed an advantage over local lateral inhibition.

Conclusions:

• Competitive mechanisms are vital for robust early vision models.
• The type of competition mechanism significantly impacts robustness, with global feedback inhibition proving more effective.