Return to agenda

This session will be broken into two 15-minute discussions.

11:30-11:45 The first will focus on the following slides, and will be led by Julius Guccione and Denise Kirschner: 

11:45-12:00 The second discussion, run by Xiaobo Zhou, Zhiwei Ji, Hua Tan, and Guy Genin, will focus on machine learning in MSM. The key questions for discussion that will be posed are the following:

Central question: How can machine learning be further harnessed for MSM prediction and validation?

Subsidiary questions:

Can machine learning networks be used to quantify biological processes?

• Omics data?
• Pathway modeling?
• Cell-cell interactions?
• Meso-scale phenomena?

Can machine learning methods be harnessed for parameter estimation/model optimization?

Can machine learning reduce parameter space in experimental design?

Can machine learning reduce medical imaging requirements for validating to multiscale models?•Can underlying models be used to quantify biological processes?

The slides for this discussion follow:

All participants are invited to add thoughts, questions, and discussion points below:

from Chase Cockrell:  I think one of the problems with getting access to HPC allocations is that oftentimes, our projects are referred to as "embarrassingly parallel" and the DoE and other supercomputing institutions are not as interested in these projects.  The types of simulations that are useful to the NIH are often structured very differently than large scale nuclear structure simulations, for example.

From Gary An: An excellent critical appraisal of deep learning from Gary Marcus: https://arxiv.org/abs/1801.00631. Particularly note 3.1 and 3.2, especially how AlphaGo worked

Also from Gary An: As noted in Comment on Challenge 1+2, danger of supposition of path uniqueness if try to use ML to do causal inference that produces generalizable knowledge, particularly given limitations of training data.