Back to 2023 Agenda
Back to main Breakout Session page
2. New Mechanistic-ML Methods: PINN, transformer, neural operators, XAI, Large Language Models
MSM lead: George Karniadakis; IMAG lead: Mauricio Rangel-Gomez, Ilana Goldberg, Julia Berzhanskaya
- What are the similarities and differences between current methods; e.g. PINN, neural operator, transformer, XAI, Large Language models?
- What are the opportunities to improve these methods?
- How can these methods be applied to blood diagnostics an diseases as it pertains to the IMAG interest group AI and Machine Learning for Blood Diagnostics and Diseases | Interagency Modeling and Analysis Group (nih.gov)?
- How can these methods be applied to neuroscience?
- How can these methods be used in the IMAG initiatives discussed in Session 2.1?
- How can we use biomechanistic and bioinformatics models synenergistically?
- Graph neural networks and causal inference.
- Uncertainty quantification in neural networks.
Resources:
Articles of Interest:
[2303.12093] ChatGPT for Programming Numerical Methods (arxiv.org)
4 Ideas for Physics-Informed Neural Networks that FAILED | by Rafael Bischof | Towards Data Science
Uncertainty quantification in scientific machine learning: Methods, metrics, and comparisons
Learning nonlinear operators via DeepONet based on the universal approximation theorem of operators
A Deep convolutional neural network for classification of red blood cells in sickle cell anemia
NIH Announcements:
IMAG wiki: Models, Tools and Databases
Comment
Here is a link to a python meetup that hosted Sophia Yang from Anaconda that discussed tools to explain AI:
Also, there are tools that combine mechanistic models and machine learning / AI techniques. I personally use ensemble models that combine different types of diease models to explain diseases - each mechanistic model is explainable to a human However, when you get them together in an ensemble many interesting things happen - see here for example:
https://jacob-barhak.github.
When data arrives from many locations such multiple clinical institutions, it will be standardized differently. A good example of this is data coming into clinicaltrials.gov from many sources
It is possible to resolve this issue using modern machine learning technique . See:
There was a discussion about the need to improve on the training efficiency of large neural networks.
Recently I came across a company called Data Heroes that uses coresets to reduce training time:
It is possible to get more details from them about papers int he scientific literature
Ilana brought up the issue of biased data.
Did you even consider Evolutionary Computational algorithms for down sampling to match a specific distribution criteria?
I must mention Aaron Garret and his Inspyred library in this context
https://github.com/aarongarrett/inspyred
It is a very useful AI tool
Link: https://www.nist.gov/publications/four-principles-explainable-artificia…
Four Principles of Explainable Artificial Intelligence
Published September 29, 2021
Author(s)
P. Jonathon Phillips, Carina Hahn, Peter Fontana, Amy Yates, Kristen K. Greene, David A. Broniatowski, Mark A. Przybocki
Citation NIST Interagency/Internal Report (NISTIR) - 8312
Report Number8312
NIST Pub Series
NIST Interagency/Internal Report (NISTIR)
Pub TypeNIST Pubs
Download Paper
https://doi.org/10.6028/NIST.IR.8312