Back to Main Agenda

Back to THEME 3- Data-Driven Approaches session page

Back to Human Safety Page

Session Lead: Linda Petzold

IMAG Moderators: Steven Lee (DOE), Jerry Myers (NASA)

Breakout Session Notes:

• Introductions (name and interest):
  • Linda Petzold
  • Jerry Myers
  • Amanda Minnich
  • Mona Matar
  • Lauren McIntyre
  • Simon Rose
  • Sanjay Purushotham
  • Casey Hanley
  • Parya Aghasafari
  • Jessica Zhang
  • Ashlee Ford Versypt
  • Ahmet Erdemir
  • Amy Gryshuk
  • Julie Leonard-Duke
  • Shantenu Jha
  • Ken Wilkins

Potential Applications

* Recovery: Instantaneous injuries, crashes. Surrogate for human. LSI model (short-term), LS-DYNA finite element code/models. ML model.

* Psychiatric drugs selection.

* Chiropractics

* Spine models. Combine simulation & experiments to improve accuracy. How to improve the (simplified) model?

* Personalize a model, what drug and predict how long drug needs to be taken.

* TBI

etc, etc, etc

Challenges

* Two sources of data: Real - noisy. Simulated data - don't know good is it? How to include physics into the digital/simulated data?

* How much patient data is needed? Depends on the model. What's the right balance?  Need enough of the right kinds of data.

* Model design and uncertainty propagation. Integrated quantities - use model for quantity of interest. Jerry model - crystals of urine. Representation of next-level of physics (multi-scale).

* What machine learning architecture to use: layer, nodes per layer, connectivity, etc? What question are you asking? See Rule 1 - Context.

* What can we do better? Can multi-modal data be combined well?

* Does the data suit the purpose? How to fuse the data?

* Guard against use ML as a hammer (need a screw driver). 

* Multi-scale processes  - can't measure what we need. Need guidance in how to interpret results/architect the framework

* How to make models credible? Best data? How to validate? What are the uncertainties and how propagated? How to document the linkages? What level of confidence in the result? 

* Probabilistic models (emphasize uncertainty, sensitivity). Advantages of Bayesian methods? What to do with small data sets?

* Not modeling for model's sake. Need testable hypothesis and check the model's accuracy.

* Probabilisitic Risk Assessment - not all models are testable. Some situations where we can not test directly.

* Lots of potential for ML: What to test next? Can the model help with that?

Obstacles

* Linking of model and data

* Hidden biases in the data

* Workflow: Data collection, Model updating, Does it converge?

* Our credibility

* Lack of expertise & background, programming skills. Educating the workforce. Need for intensive, focused workshops?

* Risk and consequences of mistakes can be high. Who is liable?

* Too much specialization? Image processing vs natural language processing, etc.  Matlab vs Python vs Julia

* Confusion about nature, scope, limitations of AI/ML. Need to manage expectations. Over-confidence.

* Need to understand the application (domain knowledge).

* Education: Apprenticeship - learning on the job. Mentors. 

* Need to have stakeholders in the discussion. What's the right level of interaction from end-users (user expectations, need specifics or broader concerns)?