時間：2026/03/06 14:00~14:20
地點：電1F01
演講者：許之凡 博士
Abstact:
Recent advances in AI, such as large language models and foundation models, have achieved remarkable benchmark performance across various tasks. However, translating these successes into high-stakes industrial environments remains a non-trivial task. In manufacturing settings, intelligence alone is insufficient; reliability, safety, robustness, and deterministic behavior are equally critical alongside raw accuracy. This talk examines the fundamental gap between research-grade AI and industrial-grade deployment. Through real-world case studies and empirical observations from manufacturing systems, we identify three key barriers to adoption: (1) the need for domain-specific adaptation beyond generic pre-trained models, (2) the challenge of establishing measurable trust in black-box AI systems, and (3) the practical friction of integrating modern AI pipelines into legacy shop-floor infrastructures. We further show representative failure modes where state-of-the-art models break down under simple constraints, and discuss the engineering trade-offs required to maintain stable and continuous production. Finally, we outline a forward-looking research roadmap centered on three pillars: automated data quality synthesis, privacy-preserving learning for proprietary data, and human-in-the-loop hybrid systems that balance automation with operational control. This talk provides empirical insights and lessons learned from real-world manufacturing deployments, offering a grounded perspective for designing AI systems that are truly production-ready.
