Evaluating the effectiveness of Deep Learning Models for Foundational Program Analysis(…)
summary
Compares CuBERT, CodeBERT, GGNN and Graph Sandwiches on two concrete tasks—alias prediction (must/may/must-not) and semantic equivalence—using a new CodeSem dataset drawn from Linux, GCC and MySQL and manually validated; CuBERT reaches 89% on alias and 84% on equivalence. Introduces a generalized + specialized pre-training pipeline that improves fine-tuning, finds token-sequence models often beat graph models, PDG boosts graph accuracy, and releases code and data for reuse.
AI-written summary · about
chapters (8)
- 0:00Deep learning for program analysis
- 1:28Evaluating deep learning for static analysis
- 3:59Training foundational program analysis
- 5:45Introducing CodeSame dataset
- 7:53Model selection and results
- 10:45Key findings and limitations
- 12:54Conclusion and future work
- 13:39Q&A
related talks
44mBuilding Knowledgeable Machines
01h 16mBoot camp on graph foundations models
01h 03mTowards Foundation Models for Graphs
40mPanel discussion "NLP"
01h 16mGeneralization in natural language processing
37mMachine Learning on Source Code
46mWhy data-driven methods will shape the future (...)
14mNeural Reverse Engineering of Stripped Binaries using Augmented Control Flow Graphs