Breakout 2 - New Mechanistic-ML Methods

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:

Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead

[2303.12093] ChatGPT for Programming Numerical Methods (arxiv.org)

4 Ideas for Physics-Informed Neural Networks that FAILED | by Rafael Bischof | Towards Data Science

10 Exciting Project Ideas Using Large Language Models (LLMs) for Your Portfolio | by Leonie Monigatti | May, 2023 | 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:

NIMH » Explainable Artificial Intelligence for Decoding and Modulating Behaviorally-Activated Brain Circuits (nih.gov)

NOT-MH-23-110: Notice of Special Interest (NOSI): Explainable Artificial Intelligence for Decoding and Modulating Neural Circuit Activity Linked to Behavior (nih.gov)

May 31, 2023: NIH Pragmatic Trials Collaboratory Announces Virtual Workshop on Getting the Right Evidence to Decision-Makers Faster - Rethinking Clinical Trials

IMAG wiki: Models, Tools and Databases

MSM Theoretical and Computational Methods WG

2019 IMAG ML-MSM Meeting

Publication from 2019 ML-MSM meeting

Comment

Comment

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

Submitted by sanjaypurushotham on Thu, 06/29/2023 - 13:00

Comment

Here is a link to a python meetup that hosted Sophia Yang from Anaconda that discussed tools to explain AI:

https://youtu.be/7NXUgiqs2jE

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.io/COVID19_Ensemble_Latest.html

 

Submitted by jbarhak on Thu, 06/29/2023 - 13:16

Comment

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: 

https://jacob-barhak.github.io/Unit_Mapping_Latest.html

Submitted by jbarhak on Thu, 06/29/2023 - 13:32

Comment

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:

https://dataheroes.ai/

It is possible to get more details from them about papers int he scientific literature 

Submitted by jbarhak on Thu, 06/29/2023 - 13:39

Comment

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

Submitted by jbarhak on Thu, 06/29/2023 - 14:04

Comment

Submitted by IlanaGoldberg on Fri, 07/07/2023 - 13:55

Add new comment