Double-correct Prediction in Sciences

A trustworthy toolbox for double-correct predictive modeling in sciences.

Abstract: AI and ML have driven scientific advances in critical domains like climate change and extreme weather prediction, but challenges remain due to unpredictable data shifts and unseen variables. This project introduces a novel trustworthy toolbox prioritizing both prediction robustness and rationale validity, ensuring accurate outcomes are backed by scientifically grounded rationales, even with unforeseen data variations. The toolbox will be optimized for scalability on HPC and released as open-source software, benefiting researchers in Earth, Marine, and Environmental Sciences through accessible, generalizable workflows and AI-ready datasets.

Publications:

• [ICML'25] Fengchun Qiao, Yanlin Chen and Xi Peng. Structure-informed Risk Minimization for Robust Ensemble Learning. In International Conference on Machine Learning, 2025. [PDF] [Code]
• [ICML'25] Mengmeng Ma, Tang Li, Yunxiang Peng, Lu Lin, Volkan Beylergil, Binsheng Zhao, Oguz Akin, Xi Peng. “Why Is There a Tumor?”: Tell Me the Reason, Show Me the Evidence. In International Conference on Machine Learning, 2025. [PDF] [Code] 
• [AAAI'25] Qitong Wang, Tang Li, Kien X. Nguyen, Xi Peng. Beyond Accuracy: On the Effects of Fine-tuning Towards Vision-Language Model's Prediction Rationality. In Proceedings of Association for the Advancement of Artificial Intelligence (AAAI), 2025. [PDF] [Code]
• [NeurIPS'24] Kien X. Nguyen, Arthur Trembanis, Xi Peng. SeafloorAI: A Large-scale Vision-Language Dataset for Seafloor Geological Survey. In Proceedings of Advances in Neural Information Processing Systems Datasets and Benchmarks Track, 2024. [PDF] [Dataset]
• [NeurIPS'24] Tang Li, Mengmeng Ma, Xi Peng. Beyond Accuracy: Ensuring Correct Predictions with Correct Rationales In Proceedings of Advances in Neural Information Processing Systems, 2024. [PDF] [Code]

Open-Sourced Data:

• Prediction Rationale Dataset for ImageNet: We construct a new rationale dataset that covers all 1,000 categories in the ImageNet. For each category, we generate an ontology tree with a maximum height of two. Combining attributes and sub-attributes, this dataset contains over 4,000 unique rationales. [https://github.com/deep-real/DCP/tree/main/Rationale%20Dataset]
• Prostate MRI and PIRAIDl Dataset: We curate a prediction rationale dataset to tackle the lack of paired MRI images, annotations, and textual rationales for AI training. The dataset includes 180K image-mask-rationale triples with quality evaluated by expert radiologists. [https://github.com/deep-real/MedRationale].

Open-Sourced Software:

• SRM Toolkit: Ensemble-based structured risk minimization for robust prediction under distribution shifts [https://github.com/deep-real/SRM]
• DCP Toolkit (Natural Images): Dual-correct training with rationale supervision for dual-correct vision models [https://github.com/deep-real/DCP]
• MedRationale Toolkit (Medical Images): Self-supervised clinical concept localization for interpretable medical image diagnosis [https://github.com/deep-real/MedRationale]