Quality Alchemist (品質煉金術師): PolyU COMP Seminar on Trends & Opportunities in AI and LFM

2024年10月10日星期四

PolyU COMP Seminar on Trends & Opportunities in AI and LFM

PolyU COMP 50^th Anniversary Distinguished Talk entitled “Trends and Opportunities in AI and Large Foundation Models” organized by Department of Computing, PolyU on 10^th Oct 2024. Prof. CHUA Tat-Seng (Chair Professor, School of Computing, National University of Singapore) was the guest speaker.

In the beginning, Prof. Chua briefed the history of AI since 1950. With the recent advancements in machine learning, particular in the emergence of large Large Foundation Model (LFM) would be the future direction. He introduced some great success of AI past case such as Deep Net, IBM’s Watson, AlphaGo, to LLMs.

And then Prof. Chua said LLMs with AGI attributes including domain versatility, output diversity, human-level semantic coherence and generative capability but some challenges we faced including hallucination, trust, safety and high cost. He also mentioned from LLMs to LFMs that multimodal tasks are current approach.

After that he discussed the evolution of smaller-sized LFMs and towards LFM-based agents to perform most functions that humans done. Prof. Chua briefed the key research directions of multimodal LFMs including multimodal alignment, alignment towards human value with trust & safety, multimodal agent, AI governance and policies, etc.

Finally, he introduced NExT-GPT’s human-machine interactions in a multimodal world.

In order to avoid hallucination in visual instruction data, simple solution is to learn to see and analyze before answering that LFMs can develop manipulation ability from prior training (Grounding, OCR, Counting) and imitate human-like behaviors (Zoom in).

Mixture of AI Experts (MoAI) concept was introduced that employed multiple Cross-validation (CV) models as experts. And then he also introduced reinforcement learning with human and AI feedback.

Lastly, Prof. CHUA Tat-Seng discussed from LFM-based agents to Network of Experts (NoE) overall architecture. In addition, “Quality assessment of LFM output is a neglected area of research but important towards self-reflective LFM” he said. At the end, he pointed out some key research topics such as LFM & Human brain relationship, super-alignment, leveraging human-machine intelligence and NoE framework.