Generative AI are revolutionary technologies with profound impacts on our daily human-computing interactions, facilitating the creation of content that emulates human creativity. Among these, Large language models (LLM), such as OpenAI’s Generative Pre-trained Transformer (GPT) foundation models and Google's Bidirectional Encoder Representations from Transformers (BERT), have become a significant focus in the current landscape. Beyond understanding their impact and limitations, there is a crucial need to ensure their correctness, verify their outputs, and promote safe usage, especially in the context of the DOE's scientific mission. It is imperative that these models are responsibly used and their broader implications are thoroughly examined to effectively serve in scientific research and exploration.

Aims and Scope

The Generative AI for ORNL Science Workshop series, invites the scientific community to share current challenges, requirements and opportunities for the safe use of generative AI technologies in our mission. Our goal is to provide a venue to educate and exchange research and development ideas, collaborations and investments around the current state-of-the-art in these relatively new technologies. We welcome lightning talk proposals from the wide range of experimental, observational, high-performance computing (HPC) projects at ORNL. We will summarize our ideas, findings and key opportunities in a subsequent report that we will share with the community and interested participants.

Registration

While this is a hybrid event, we encourage in-person participation to improve interactions with the speakers and within the community.

Registration Link: Please register by July 13, 2023 (CLOSED).

External (virtual or in-person) participants need to register. If you are attending SMC23 (https://smc.ornl.gov), you are automatically approved to attend but you still need to register with a special registration link. ORNL (virtual or in-person) participants need to register with an internal registration link provided by the workshop organizers.

Note: Registration is open until July 13 for external and non-SMC23 attendees.

The registration form includes the option to propose a talk focusing on the requirements for scientific areas that are representative.

• Registration is required for in-person or virtual participation via Zoom
• We encourage early registration for in-person participation due to the venue’s limited capacity

Call for Talks

We invite participants to present a talk that aligns with the scope of the workshop (option provided in the registration form) Full talks (45mins) and lightning talks (2-3mins) are encouraged.

Topics:

• Requirements and conditions for applying Generative AI in scientific contexts
• Implementing safeguards and verification methods for generative AI models to ensure safety and correctness
• Exploration of Large Language Models (LLMs): delving into models such as GPT, Bard, and more
• Ensuring safe usage of generative AI in observational, experimental, and computational science: the balance between innovation and corretness
• Assessing the transformative impact of LLMs on the scientific discovery process: the advantages, potential limitations, and broader implications

Agenda

Location: Tennessee Ballroom, at the Crowne Plaza Hotel Downtown, Knoxville, TN

Time	Session	Presenter
8:00am-8:30am	Registration
Working Breakfast
8:30am-9:15am	Opening Remarks and Keynote: AI Security Research - A new science at the intersection of AI and cybersecurity research	Prasanna Balaprakash, AI Initiative Director and Edmon Begoli, AI Systems R&D Section Head, ORNL
9:15am-9:45am	DecodingTrust: Assessing Trustworthiness and Risks of Generative Models	Bo Li, Neubauer Associate professor, University of Chicago
9:45am-11:00am	Lightning Talks Session	Session Chair: William F Godoy, ORNL
	Toward Multimodal Foundation Models for GeoAI	Philipe Ambrozio Dias, ORNL
	Toward a Foundation Model for Climate Research and Applications	Valentine Anantharaj, ORNL
	How Generative Deep learning could change biology	Ada Sedova, ORNL
	TwoFold at Scale	Jens Glaser, ORNL
	AI at NERSC	Wahid Bhimji, LBNL/NERSC
	Developing deep generative models for scientific data	Jong Choi, ORNL
	AutoMin: Learnings from Generative NLP Models for Automatic Generation and Evaluation of Meeting Minutes	Tirthankar Ghosal, ORNL
	Blackout Diffusion: Generative Models in Discrete State Spaces	Zachary Fox, ORNL
	Trustworthy graph neural networks using new hardware for energy efficiency	Massimiliano Lupo Pasini, ORNL
	FORGE: Open Foundation Models for Science	Junqi Yin, ORNL
11:00am-11:10am	Break
11:10am-11:15pm	Introduction to the Vendor’s Session	Session Chair: Oscar Hernandez, ORNL
11:15am-12:00pm	From pre-training to model alignment	Sandeep Subramanian, NVIDIA
Working Lunch
12:00pm-12:45pm	Generative AI and Large Language Models with SambaNova Systems	Jennifer Glore and Ken Kutzer, SambaNova
Afternoon
12:45pm-1:00pm	Break
1:00pm-1:45pm	Low-Latency Inference at Scale in the age of LLMs and ML Accelerators	Andrew Ling, Groq
1:45pm-2:30pm	Training LLMs with Cerebras	Richard Kuzma, Cerebras
2:30pm-3:15pm	Efficient Generative AI with Graphcore IPUs	Chad Martin, Graphcore
3:15pm-3:30pm	Closing remarks, adjourn

Organizers

• Prasanna Balaprakash
• Oscar Hernandez
• William F Godoy

Program Commmitee

• Prasanna Balaprakash
• John Gounley
• Aris Tsaris
• Isaac Lyngaas
• Mayanka Chandra Shekar
• Jens Glaser
• Junqi Yin
• Ada Sedova
• Edmon Begoli
• Amir Sadovnik
• Dalton Lunga

Logistics and Planning Chair

• Taylor Bullock

Steering Committee

• Andrea Delgado
• Keita Teranishi
• Pedro Valero-Lara
• Jeffrey S Vetter

Hotel Information

• SMC’23 hotel information

Sponsors:

• Computing and Computational Sciences Directorate
• ASCR Bluestone Project
• ORNL AI initiative

The workshop organizers greatly appreciate the support of our corporate sponsors: