Back to Main Agenda

Back to THEME 2- Partial Differential Equations session page

Session Lead: George Karniadakis

IMAG Moderators:  Fariba Fahroo (AFOSR), Tony Kirilusha (NIAMS)

Breakout Session Notes:

• Introductions (name and interest):
  • Session 1:
  • Session 2:
• Build on current state of the art PDE Models
  • Leverage the sources of certainty (physics) to mitigate sources of uncertainty (data)
  • Use ensembles of PDEs
  • Node-2-Vec – allows to find effective dimensionality of a system but letting dimensions with a very high uncertainty fall out.
  • How high can we include uncertainty propagation into “towers” that comprise multiscale models?
• Build on current state of the art models for Digital Twins
  • Models to study human movement go into the hands of physicians, who just want it to work
  • Train the model to be adaptable to a specific patient (not a population average) – this is a very difficult problem.  Pose the problem by assuming the parameters are a sample from a characterized population, fit measurements from a new patient onto existing parameters (which measurements will give you the maximum amount of information).
  • Multiple layers of anatomy in the digital twin – what are the opportunities for “persistent training” (continuous learning over time)?
  • Do you need to simulate data for “rare” phenotypes (rare diseases vs. common ones)? – does it address the question of cohort heterogeneity in cases of “common disease”.
  • Data access and data privacy?  Harder to get human data than non-human data.
  • Can you explain why some drugs work in animals but fail in humans (or the other way around?)
  • How does instantaneous injury develop into a chronic injury?
• ML-MSM integration opportunities
  •  
• Challenges ML-MSM modelers should address
  • Where does your PDE come from?  If a PDE represents folding tissue, how do you pick the right one?  Do you trust the PDE?
  • Rich field of uncertainty quantification.  There is a lack of work that addresses confidence in machine-learning (ML) predictions.  These ideas from the PDE world should be transplanted to ML?
  • Reproducible research computing - disseminating scripts and code, standardized language to describe models.
  • Model unceratinty in the context of ML is dependent on the training data (and on the network itself).  What uncertainties are there?  (Parametric uncertainty?)
  • In addition to the model, also model the measuring process?  Data come from different modaities?  What is the hierarchy of data - which ones do you trust more?
  • Should you model the data collection process?  Is that feasible?
  • Brittleness of models is an ongoing issue.  Neural networks do better if exposed to noise during training.
  • The idea is to make predictions out of the scope of the orinigal data.  At the end of the day - does it work?
  • Integrating prior knowledge of physics into the model to improve performance.
  • Different architectures of neural networks are more appropriate to certain types of data than to others.
  • Potential barrier to standardized adversarial training is speed - adversarial training is very computationally expensive.
  • Would an ensemble of PDEs work well?  Model merging / convergence.
  • Is it possible to combine human and animal data to improve overall performance of a model for human predictions.  For critical tasks models should come back with a confidence parameter.