Optimal Mixture-of-Experts Model Averaging for Conditional Generative Learning
Holder： Baihua He（University of Science and Technology of China）
Time：2026-04-23 15:10-17:00
Location：Room 217, Guanghua Building 2
Abstract:
Learning complex conditioned distributions is essential in modern generative modeling, powering applications from probabilistic prediction to AI-driven image and text synthesis. Different generative models often excel on some conditioning inputs but struggle on others, yet classical ensemble methods assign fixed weights that ignore this variability. We introduce a statistically principled Optimal Mixture-of-Experts framework that learns input-dependent weights to combine multiple conditional generators dynamically. By extending model averaging theory and using integral probability metrics to align distributions, we prove our adaptive weighting achieves asymptotic optimality under broad conditions. The proposed method can seamlessly incorporate any conditional generator and can be extended to other probabilistic tasks. Comprehensive simulations and real-world experiments demonstrate that our method consistently outperforms individual models and traditional ensembles. These experiments, such as image generation and demand distribution learning, show its broad effectiveness in other statisitic and AI applications.
About the Speaker:
贺百花,中国科学技术大学特任副教授,主要从事模型平均以及多源域学习等领域的研究,相关论文发表在JASA,JMLR,IJOC等统计和机器学习期刊,主持青年基金以及参与国家自然科学基金重大项目和科技部重点研发计划,入选省部级人才计划支持.
Your participation is warmly welcomed!
欢迎扫码关注北大统计科学中心公众号，了解更多讲座信息!
