The aim of the workshop is to explore what it means to build “cross-experiment multi-modal” foundation models in a landscape where the relevant scientific data can be subject to a range of policies from “open data” to experiment-restricted (either proprietary or limited distribution/raw) data as well as the use of simulations in this context. Similarly, both the data and simulation come with a range of latency/embargo, “initial use vs reuse/reinterpretation” and experiment governance structures. The goal of this workshop is to explore the technical, policy and cyberinfrastructure questions that arise when pursuing such shared models.
Specific questions:
-
What does it actually mean in practice to build a “foundation model” across experiments with different detector designs, data formats, and physics goals?
-
What does it mean in practice to do pre-training on “diverse data” in a shared environment vs fine-tuning in a restricted (experiment) environment?
-
How is benchmarking of the models done? How are the models validated when issues arise spanning the pre-training and restricted fine-tuning?
-
What are the technical and cyberinfrastructure implications?
-
Who owns the resulting models and what are the implications given different experiment governance structures? (And data ownership by international collaborations?)
-
If initially trained on a set of current and archived data, how do these models evolve going forward as new data appears from new experiments/upgrades/detector configurations?
-
If industry is involved in parts of this process, how do we avoid issues related to vendor lock-in and/or retain the “public” expectation that underlies most of the government funding of fundamental science?
 |
This event is sponsored in part by the National Science Foundation through grants OAC-2226378, OAC-2226379 and OAC-2226380 (FAIROS-HEP) Any opinions, findings, conclusions or recommendations expressed in this material are those of the developers and do not necessarily reflect the views of the National Science Foundation. |