IET Awards ceremony cum talk

IET Awards ceremony cum talk by Prof. Hongxia Yang, Associate Dean of the Faculty of Computing and Mathematical Sciences and Executive Director of the PolyU AI (PAAI).

Prof. Hongxia Yang introduce a new framework that makes foundation model development more efficient, accessible, and collaborative. At its core is Domain-Adaptive Continual Pretraining (DACP), which combines lightweight efficiency—via the first end-to-end FP8 low-bit training suite—with collaborative expertise, embedding professional intuition through domain-specific models. DACP continuously adapts Large Language Models with unlabeled domain data, achieving superior specialization in underrepresented fields such as healthcare and science, while reducing GPU costs and outperforming mainstream systems. The framework also features an Advanced Model Fusion module. By integrating top-performing domain models regardless of their pretraining architecture, it constructs powerful foundation models in just 160 GPU hours, compared to the million-plus hours typically needed from scratch. Its Resource-Efficient Architecture further democratizes AI by enabling effective use of distributed, entry-level GPU clusters, lessening reliance on centralized compute. We also establish the first theoretical scaling law of model merging, providing a new principle for distributed generative AI. Real-world applications demonstrate its impact, with medical models surpassing Google MedGemma and multi-agent systems rivaling OpenAI’s Deep Research. This framework reimagines foundation model development—faster, cheaper, and more inclusive.