Stochastic Concept Bottleneck Model Reproduction and Optimization

Research project / reproduction study.

Collaborators: Xingqiao Hu, Yiran Wang, Kangyue Xu.

• Reproduced CBM, CEM, Autoregressive CBM, and SCBM on synthetic, CUB, and CIFAR-10 benchmarks, analyzing concept prediction, target accuracy, calibration, and intervention behavior.
• Refactored the original research codebase into modular training and intervention pipelines, and resolved key performance bottlenecks through vectorized concept modules, batched Monte Carlo sampling, optimized data loading, and GPU-accelerated SCBM intervention solvers.
• Reduced estimated full experimental runtime from 1,200 hours to around 60 hours while preserving experimental semantics, making future research on CBMs much more accessible and efficient.

Links

• Code