Finding Bugs in Deep Learning Programs
summary
Two concrete systems are described: a fast static analysis tool NeuralLens that builds a meta‑model and applies a taxonomy of typical faults and a dynamic inspector that instruments training to extract signals. The static checks find parameter, activation and API misuses; the dynamic checks detect vanishing gradients, unstable training and overfitting. NeuralLens runs quickly; the instrumentation adds small overhead. Empirical claim is specific: the dynamic approach catches about 30 more bugs than sagemaker.
AI-written summary · about
chapters (6)
- 0:00Introduction
- 1:10Deep Learning Pipeline
- 1:45Deep Learning Failures
- 2:40Approach
- 3:40Workflow
- 4:20Dynamic Analysis
related talks
24mTesting Techniques for Critical Deep Learning Applications
17mTraining with Confidence: Catching Silent Errors in Deep Learning Training with...
25mDeepBugs - A Learning Approach to Name-based Bug Detection
48mUsing Machine Learning to find Linux bugs
58mSecurity Static Analysis - Avoiding an Angry Mob of Engineers
43mDebugging AI
18mPreview: Verification (Correctness & Debugging)
18mEnhancing Static Analysis for Practical Bug Detection: An LLM-Integrated Approach