Evaluation-Strategy Gap in Fault Diagnosis of Deep Learning Programs

2026-06-24 · Source: cs.SE updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Expert, medium

Summary

A study on fault diagnosis in Deep Learning (DL) programs reveals a significant "evaluation-strategy gap" when assessing diagnostic techniques. Researchers investigated this gap using DynFault, a corpus of 5,542 fault-injected training traces from 38 real-world DL programs. They found a 0.190 balanced accuracy drop for existing fault diagnosis techniques when moving from within-program cross-validation to program-held-out evaluation, where programs in the test set are entirely unseen during training. This performance decline is attributed to program-level structural features. The analysis further showed that curvature features are effective for detecting instability in unseen programs, while optimizer and activation features primarily benefit diagnosis only on programs present in the training data, highlighting a limitation in their generalizability.

Key takeaway

For MLOps Engineers evaluating or deploying deep learning fault diagnosis tools, you must adopt program-held-out evaluation strategies to accurately assess real-world performance. Relying solely on within-program metrics will overestimate diagnostic capabilities, especially for unseen programs. Prioritize solutions that demonstrate generalizability, such as those leveraging curvature features for instability detection, and critically re-evaluate the cross-program utility of optimizer and activation features before investing in their logging.

Key insights

The evaluation strategy significantly impacts DL fault diagnosis performance, revealing a gap in technique generalizability.

Principles

• Fault diagnosis performance is overestimated by within-program evaluation.
• Program-level feature structure can hinder diagnostic generalization.
• Runtime feature effectiveness varies significantly across evaluation strategies.

Method

The study used DynFault, a corpus of 5,542 fault-injected training traces from 38 DL programs, comparing within-program and program-held-out evaluation strategies across three diagnostic tasks and two feature sets.

In practice

• Prioritize program-held-out evaluation for new DL fault diagnosis techniques.
• Integrate curvature features for robust instability detection.
• Re-evaluate optimizer/activation features for cross-program utility.

Topics

• Deep Learning Fault Diagnosis
• Evaluation Strategy Gap
• DynFault Corpus
• Curvature Features
• Training Instability
• Model Generalization

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, MLOps Engineer

Related on AIssential

• Understanding Generalization through Decision Pattern Shift — Generalization failure in DNNs is a systematic drift in internal decision patterns, quantifiable by Decision Pattern Shift (DPS).
• Evaluating Supervised Machine Learning Models: Principles, Pitfalls, and Metric Selection — Effective model evaluation requires aligning metrics and validation with operational objectives, avoiding common pitfalls.
• Leveraging systems' non-linearity to tackle the scarcity of data in the design of Intelligent Fault Diagnosis Systems — Leveraging system non-linearity and novel visualization creates data for DTL-based Intelligent Fault Diagnosis Systems despite scarcity.
• The Model Already Knows Where It’s Wrong — A GBDT's internal geometry reveals local weaknesses, enabling diagnosis-specific model repair beyond global metrics.
• Questioning the Coverage-Length Metric in Conformal Prediction: When Shorter Intervals Are Not Better — Standard Conformal Prediction metrics can be deceptively improved by methods introducing randomness, necessitating new evaluation criteria.
• Demystifying Residual Analysis: A Beginner’s Guide to What Your Model Isn’t Telling You — Residual analysis diagnoses model failures by examining prediction errors for hidden patterns and assumption violations.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.SE updates on arXiv.org.