Beyond Distribution Shift: Spurious Features Through the Lens of Training Dynamics

Area of Science:

• Machine Learning
• Deep Learning
• Artificial Intelligence

Background:

• Deep Neural Networks (DNNs) often learn spurious features that correlate with labels but are irrelevant, hindering generalization and safety-critical applications.
• Previous research focused on the negative impact of spurious features, but the distinction between benign and harmful spurious features based on learnability is underexplored.

Purpose of the Study:

• To investigate the learning dynamics of internal neurons in DNNs to better understand the effects of spurious features.
• To differentiate between benign and harmful spurious features based on their 'easiness' or 'hardness' relative to core features.

Main Methods:

• Utilized instance difficulty methods, specifically Prediction Depth, to quantify feature 'easiness' during DNN training.
• Analyzed the learning dynamics of early layers in DNNs to detect the emergence of harmful spurious features.
• Validated findings on diverse medical and vision datasets (simulated and real) and connected Prediction Depth to information-theoretic concepts.

Main Results:

• Demonstrated that not all spurious features are harmful; their impact depends on whether they are easier or harder to learn than core features.
• Identified that harmful spurious features can be detected by observing the learning dynamics in the early layers of DNNs.
• Showed that accuracy monitoring alone is insufficient for detecting spurious features during training.

Conclusions:

• Harmful spurious features can be identified by monitoring the 'easiness' of features learned by early DNN layers during training.
• Instance difficulty metrics are crucial for detecting spurious features, complementing traditional accuracy monitoring.
• Understanding and monitoring early training dynamics are essential for improving DNN generalization and reliability.