Event Details

The remarkable performance of large language models (LLMs) masks a critical challenge: a lack of understanding about their underlying mechanisms. This talk explores our work on neuron-level guidance, shifting the focus from prompt engineering. Key findings include: (1) the identification of distinct neuronal circuits for memorization and generalization; (2) the discovery of role-encoding neurons, enabling performance exceeding expert levels; and (3) a data-centric training strategy that improves LLM learning efficiency.

Join Zoom Meeting
https://usc.zoom.us/j/94409584905?pwd=Sm5LVkd0bndUdEluM3piK0NWTUQrUT09

Meeting ID: 944 0958 4905
Passcode: 822247

Host: Fred Morstatter & Abel Salinas

POC: Justina Gilleland

Speaker Bio

Prof. Shou-De Lin is a leading AI and machine learning researcher and innovator with 25 years of experience in the field. Known for his commitment to practical machine learning, he expertly translates cutting-edge theory into impactful real-world solutions. Currently a professor

at National Taiwan University, he previously led AI innovation as Chief Scientist at Appier. His diverse expertise encompasses recommendation systems, generative models, and edge-based AI solutions, realized at OmniEyes, a company he co-founded. Prof. Lin's contributions have been widely recognized through awards such as multiple ACM KDD Cup wins, AOARD grant research awards, and his designation as an ACM Distinguished Member, alongside research grant awards from top AI companies like Nvidia, Google, and Microsoft.