An AI manager’s journey into to the new era of Foundation Models.
NASA SMD AI WORKSHOP 2024
trillium.tech
James Parr
Director FDL.ai
CEO, Trillium Technologies
As an AI manager, what do you need to think about?
My job: understand the journey we need to go on before we get Foundation Models that a space agency can release.
Satellite Data is much harder to work with than text.
For science it is useful to build systems that can learn from multiple modalities.
CORE CONCEPT 1
CORE CONCEPT 2
Science is working in dynamic settings.�It’s more complicated than anyone thought.
CORE CONCEPT 3
The dream has been:
“We don’t need the physics, just let SSL do the rest”
So we talk about Foundation models as being ‘generalizable’. - e.g. in use-case (temporal or spatial)
SAR Foundation Model precursor extracts vegetation cover from SAR with just 1% of the labels needed for supervised methods.
Best Paper NeurIPS Climatechange.ai
In SAR-FM, the breakthrough is generalizability across AOIs
Polarization
Seasonality
ChatGPT doesn’t have this problem.
(It’s more complicated than we thought.)
CORE CONCEPT 3
Science data is
statistically dynamic…
“Robustness”
We need RECIPES.
CORE CONCEPT 4
Can you identify the locations first where your Foundation model will work?
Version 1
Version 2
Version 3
How can we still make Foundation models scientifically useful at small scales?
It’s a phased approach that is all about trade-offs
The initial models will need task specific recipes
Models will also have to be designed on context of adapters, integrations and narrow applications.
A glimpse of the future…
Think about robustness to inputs
The Recipe can help users hit the ground running and get useful results from precursor efforts
Physics can help!
Usefulness
Robustness
CORE CONCEPT 1
CORE CONCEPT 2
CORE CONCEPT 3
Closing thoughts and an opportunity!
Thank you!
James Parr
Director FDL.ai
CEO, Trillium Technologies
james@fdl.ai