Title: Large Language Models: What's Happening Now?

Speaker: Prof. Keisuke Sakaguchi @ Tohoku University

Time : 2023 April 5th 14:00 ~ 15:00

Location: Online

Online: https://us02web.zoom.us/j/87084697053?pwd=U1ZqUXpJMHczcW1pKzJXck01NHZhUT09 (Passcode: 0405)

Language: English speech & English slides

Abstract: 

Large Language Models (LLMs) such as BERT, GPT-3, and the recently introduced GPT-4 have ushered in a new era in natural language processing, enabling applications like machine translation, summarization, proofreading, and chatbot dialogues across diverse industries and start-ups. Powered by the Encoder-Decoder (Transformer) architecture and self-attention mechanisms, LLMs harness vast amounts of web data to achieve remarkable results. GPT-4 has introduced advancements in instruction fine-tuning and reinforcement learning with human feedback (RLHF), further enhancing its capabilities. In this presentation, we will explore the current landscape of LLMs, investigating their inner workings, recent advancements, and the implementation of these methods. Additionally, we will discuss their limitations and identify emerging research directions, incorporating insights from our recent projects in the field. (By the way, a portion of this abstract was crafted by GPT-4. Can you identify which part?)

Bio:

Keisuke Sakaguchi is an associate professor at Tohoku University. His research interests lie at the intersection of Natural Language Processing, Machine Learning, and Psycholinguistics. The long-term research goals are 1) to understand human intelligence, especially natural langauge processing, and 2) to build embodied AI that is as robust and efficient as humans. More specifically, his research interests include: robust NLP models for noisy texts (e.g., text normalization, parsing, automated grammatical error correction), commonsense knowledge acquisition and reasoning, NLP for educational purposes (first and second language acquisition), text generation (decoding algorithms), efficient data collection, and reliable evaluation metric design (meta-evaluation).