The intriguing role of module criticality in the generalization of deep networks
This work addresses a fundamental issue in deep learning theory for researchers and practitioners, but it appears incremental as it builds on existing concepts of loss landscape analysis.
The paper tackles the problem of understanding why some modules in deep neural networks are more critical for performance than others, and proposes a complexity measure called module criticality that explains generalization differences between architectures, showing it outperforms earlier measures.
We study the phenomenon that some modules of deep neural networks (DNNs) are more critical than others. Meaning that rewinding their parameter values back to initialization, while keeping other modules fixed at the trained parameters, results in a large drop in the network's performance. Our analysis reveals interesting properties of the loss landscape which leads us to propose a complexity measure, called module criticality, based on the shape of the valleys that connects the initial and final values of the module parameters. We formulate how generalization relates to the module criticality, and show that this measure is able to explain the superior generalization performance of some architectures over others, whereas earlier measures fail to do so.