Learning Layer-Skippable Inference Network.

Area of Science:

• Computer Vision
• Machine Learning
• Computational Neuroscience

Background:

• Learning effective representations for machine learning is computationally intensive.
• Current inference methods differ significantly from biological visual systems.
• Human visual processing involves hierarchical, frequency-based analysis with feedback loops.

Purpose of the Study:

• To explore layer-skippable inference networks inspired by biological vision.
• To develop a dynamic network for coarse-to-fine object categorization.
• To reduce computational costs while maintaining or improving performance.

Main Methods:

• Proposed a layer-skipping network with two branches for coarse and fine-grained classification.
• Developed a gating network to generate dynamic inference graphs and enable layer skipping.
• Introduced a ranking-based loss function for efficient gating network training.
• Incorporated top-down feedback and feature-wise affine transformation for representation enhancement.

Main Results:

• Achieved promising results on coarse-to-fine object categorization benchmarks.
• Demonstrated improved network robustness against adversarial attacks.
• The layer-skipping mechanism dynamically adjusts inference paths, reducing computational load.

Conclusions:

• The proposed layer-skipping network offers an efficient and flexible approach to object categorization.
• The model's architecture, inspired by neuroscience, yields significant performance and robustness benefits.
• This adaptive inference strategy leads to high-performance deep networks with dynamic structures.