Back to Main Agenda

Back to THEME 4- Theory-Driven Approaches session page

Back to Human Safety Page

Session Lead: Paris Perdikaris

IMAG Moderators: Elizabeth Ginexi (NIH-OBSSR), Rachel Slayton (CDC)

Breakout Session Notes:

• Introductions (name and interest):
  • Session 1:
  • Session 2:
• Build on current state of the are Theory-Driven Models

  • Use prior knowledge where it exists.

    Prior knowledge comes in different forms (as data comes in different modalities):

    • • Mechanistic models of conservation laws (ODEs, PDEs, etc.)
      • Invariance to symmetry groups
      • Features selected by expert knowledge, simulated data, historical data, population data, etc.

    Effect of prior knowledge:

           -- Good generalization/extrapolation performance to new unseen cases

            -- Sample efficiency (can we train deep learning systems with no data?)

            -- Interpretability

    How to embed it?

    • Implicitly embed it in the model architecture
    • Priors/regularization/constraint of the loss function

    Caution: Embedding prior knowledge cannot be treated as a black-box. It requires us to rethink our model’s architecture, how the model is initialized, how it is optimized/trained, etc.

• Build on current state of the art models for Human Safety

  • Would you trust a neural network to be robust to small perturbations in data, e.g., pacemaker design 

    • Do you trust the physics or the neural network more? 

    • What’s the bias in the data? Can the physics better constrain it? 

    • Can you capture causal effects in observational data? 

           Showcase predictive power: randomized trails (ideally), or some form of demonstrative data 

• ML-MSM integration opportunities
  • Encode prior knowledge in the networks –saying there is certainty in the lack of connection if it is not included
  • Distilling emergence of function: from the micro-scale to the macro-scale. The role of information loss and coarse graining.
  • Commonalities and differences in learning of biological and artificial systems. Using reinforcement learning to understand evolution.
  
  
  •  
• Challenges ML-MSM modelers should address

  • Challenges: 

    • Unsatisfying to just push data through an algorithm 

    • Choices to add to data or add to knowledge -- a black-box approach is unlikely to work, need to choose the right tool for the given problem

    • Need an ability to extract mechanisms rather than just correlation; want to know what levers there are to change 

    • Success of neural networks is largely based on the architecture of the networks. Making the art of model calibration more systematic.

    • Hard to pinpoint the laws of a system and understand why it acts in that way

    • Uncertainty quantification

           Recurring feedback is poorly captured

           Need to explain causality and quantify counterfactuals

     

          Data privacy, data ownership of an individual's data, and how it may influence treatment decisions

          Ultimately it all blends down to decision making: What are the decisions to be made and who are the appropriate decision makers (e.g., physicians, patients)?