DeepBugs - A Learning Approach to Name-based Bug Detection

on YouTube ↗25m

" summary

Uses simple code transformations to seed millions of training examples, learns identifier embeddings with word2vec and concatenates those vectors into a shallow feed‑forward neural classifier to detect swapped arguments and two kinds of binary-op mistakes. Trained on ~150,000 JavaScript files (≈86M LOC), classifiers reach about 89–95% accuracy, run under 20 ms per file, and surface 102 real-world mistakes with a 68% true‑positive rate, showing seeded bugs generalize to real defects.

AI-written summary · about

" chapters (15)

1. 0:00Intro
2. 0:13Automated Bug Detection
3. 1:36Learning to Find Bugs
4. 3:46Name-related Bugs (2)
5. 4:36Overview of DeepBugs
6. 5:52Generating Training Data
7. 7:42Representing Code as Vectors
8. 8:53Example: Embeddings
9. 12:08Predicting Bugs in New Code
10. 12:38Evaluation: Setup
11. 13:40Accuracy of Classifier
12. 14:12Examples of Detected Bugs
13. 15:48Precision
14. 16:51Importance of Embeddings
15. 18:33Conclusion

" related talks

• 07mFinding Bugs in Deep Learning ProgramsFoutse Khomh · Strange Loop 2022
• 38mApplying machine learning to software development to reduce bugsNitin Borwankar · PyData San Francisco 2016
• 07mInterviewFrancesc Campoy Flores · Scale By The Bay 2018
• 48mUsing Machine Learning to find Linux bugsStef Walter · DevConf.CZ 2019
• 47mBug prediction based on your code historyAlexey Tokar · JavaDay Ukraine 2017
• 37mMachine Learning on Source CodeFrancesc Campoy Flores · Scale By The Bay 2018
• 50mPractical Applications of ML in Software TestingMesut Durukal · NDC Sydney 2020
• 30mData Driven SecurityPeter Elmer · stackconf online 2021