IEEE SMC-IT/SCC 2024: Trustworthiness of Foundation Models and What They
Generate
A mini workshop as a part of
SMC-IT/SCC 2024, July 15-19, 2024
Computer History Museum, Mountain View, CA, USA
Organizers: Daniel Crichton and Richard Doyle
(JPL/IEEE) and Ashish Mahabal (Caltech)
Image credit: Robert Hurt (Caltech/IPAC)
Motivation for the workshop:
Interest in Foundation Models (FMs) has burgeoned, with much attention focused
on Large Language Models (LLMs): their notable performance, along with open
issues such as hallucinations. LLM success appears to rest on the sheer power
of learning statistical patterns from massive public databases, and on
inherent ordering constraints within a domain, exploited by completion
operations. Yet there is growing interest in other applications where such
constraints may be less prevalent, or different in form. Within NASA, there is
interest in exploring the potential of FMs for analyzing and understanding
image and time-series data.
Trustworthiness is a key challenge, which must entail some form of validation
of the FMs themselves. Explainability is an established research challenge
relevant to uses of Artificial Intelligence (AI), particularly Machine
Learning (ML); closely related is human-machine interaction. In the context of
FMs, there is an important responsibility on users—not sufficiently
emphasized—to contribute to the validation of Generative AI outputs—if not the
FMs themselves—as one antidote to hallucinations. Our workshop will convene
SMEs in these research areas, to discuss extant challenges and possible paths
to solutions, noting—as is not unusual with AI—where existing best practices
(e.g., in V&V) may be adaptable / extensible, and where new methodologies may
be required.
One of the claims, and attractions, of FMs is their alleged zero-shot
applicability to diverse corpora of text, images, and other data types.
Validation of such claims lies at the intersection of developing FMs for
fundamental sciences and their trustworthiness. For instance, can the Segment
Anything Model (SAM) work for fuzzy images (say a cancer starting to attack an
adjacent organ and thus blurring boundaries), or an LLM trying to address
questions from a field that has sparse corpora (even when supported with
mechanisms like retrieval-augmented generation). With time-series the
zero-shot applicability will be even more under scrutiny (for instance for
gappy data as in much of terrestrial astronomy).
Workshop Structure:
The workshop will have two 90-minute tracks, each featuring three 20-minute
invited talks, followed by a a panel
of speakers, moderated by the workshop co-chairs. The SMEs will be drawn
from NASA, DoD, Academia, and Industry, including Commercial Space. The
workshop hosts will seek opportunities to cross-fertilize discussion across
both near-term and far-term considerations—with panel members, and with
workshop attendees.
