Marvel Collin / Projects
Quantization and Explainability
Quantization and Explainability is a study of what quantization does to model explanations, built by Marvel Collin for ICCSCI 2025. It was built from Feb 2026 to Jun 2026 with Python, PyTorch, IndoBERT, LIME and SHAP. Role: Solo, Research.
- Year
- Feb 2026 to Jun 2026
- Role
- Solo · Research
- Stack
- Python · PyTorch · IndoBERT · LIME · SHAP
- Client
- ICCSCI 2025
- Result
- Accepted at ICCSCI 2025
Most quantization papers report accuracy, latency and model size. This one asks a different question: after you quantize an Indonesian sentiment classifier, does it still give the same reasons for its predictions?
The pipeline fine tunes IndoBERT on the SmSA three class sentiment task from IndoNLU, then produces eight deployment variants across post training quantization and quantization aware training.
Every variant is measured on task performance, calibration error, explanation stability across LIME, SHAP, Occlusion and Integrated Gradients, and explanation faithfulness, so that a stable explanation is not just stably wrong. Stability is bootstrapped and Bonferroni corrected.
On top sit robustness stress tests, an Indonesian linguistic probe suite covering negation scope, reduplication, hedging and contrastive markers, and a deployment recommender that ranks the eight variants under configurable constraints.
Outcome: Accepted at ICCSCI 2025.