AI
Gap in Deep Learning Fault Diagnosis Revealed Across Different Programs
New research shows existing fault diagnosis techniques for deep learning drop significantly in performance when applied to unseen programs, highlighting a mismatch between evaluation methods and real-world deployment scenarios.
A new study from arXiv's software engineering track reveals a critical performance gap in fault diagnosis techniques for deep learning programs, uncovering why methods that work well in controlled settings often fail in practice.
The Problem#
Deep learning systems can fail during training for numerous reasons, and identifying the cause has traditionally been a labor-intensive debugging task. Researchers have developed various fault diagnosis techniques, but there's a disconnect between how these tools are tested and how they perform in real-world deployment.
Most existing techniques are evaluated using within-program cross-validation—testing on variations of the same programs used during development. This approach masks a critical limitation: their performance degrades significantly when applied to entirely new, previously unseen programs.
The Findings#
The research team analyzed DynFault, a corpus of 5,542 fault-injected training traces from 38 real-world deep learning programs. Their results are striking:
- Performance gap of 0.190 in balanced accuracy between within-program evaluation and holding out whole programs
- The gap stems from program-level structure embedded in the features used by diagnosis techniques
- Different feature types show markedly different behavior on unseen programs
Feature Insights#
The researchers examined two key runtime feature sets:
- Curvature features proved valuable for detecting instability problems in new, unseen programs
- Optimizer and activation features only reliably helped diagnose faults in programs the model had seen during training
This asymmetry suggests that some features capture domain-general patterns about deep learning failures, while others are overly specific to individual program characteristics.
Why It Matters#
The findings highlight a fundamental challenge in machine learning tooling: tools that achieve impressive benchmark results may not generalize to practitioners' real problems. For deep learning development environments, this means fault diagnosis systems need to be evaluated—and potentially redesigned—to handle programs outside their training distribution, not just variations of known programs.
