Corruption depth: Analysis of DNN depth for misclassification.

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Deep neural networks (DNNs) excel in computer vision but their depth requirements concerning input data complexity and noise remain unclear.
• Existing research inadequately addresses how common corruptions impact specific layers within DNNs, hindering a full understanding of classification processes.

Purpose of the Study:

• To introduce and define the concept of "corruption depth" for identifying critical layers in DNNs affected by data corruptions.
• To investigate the relationship between input data complexity, noise types, and the necessary network depth for accurate classification.
• To explore how understanding corruption depth can enhance model explainability and enable efficient model compression strategies.

Main Methods:

• Introducing the novel concept of "corruption depth" to track misclassifications through network layers.
• Conducting extensive experiments to analyze how DNNs process examples under various corruption conditions.
• Developing a methodology to identify the specific network depth at which misclassifications persist.

Main Results:

• Demonstrated that misclassification often localizes to specific network layers rather than being uniformly distributed.
• Provided insights into the processing of examples through the network, illustrating concepts of example memorization and sample difficulty.
• Quantified the impact of different corruptions on network performance at various depths.

Conclusions:

• The "corruption depth" concept offers a new perspective for understanding model behavior and diagnosing errors in deep neural networks.
• Identifying critical layers for corruption impact allows for targeted model pruning, offering a more computationally efficient alternative to full network purification.
• This approach advances model explainability beyond attention maps, enabling visualization of classification progress throughout the network and facilitating effective model compression.