Tuesday, August 20, 2019

AI meeting notes


gradient descent is robust form of optimization. Exact optimization for training data probably do not generalize well for testing data.

gradient descent 'converge' to a global optimization point?

classification,
inverse problem in materials
transportation
mobility

AI only learn what data is (so memorization?)

AI is just a marketing term?

learning: physics based, foundational math, representation learning, reinforcement learning, adversary networks,

scalability: algorithms, convergence, parallelization, mix precision arithmetic, hardware,

Assurance: uncertainty quantification, explainability and interpretability, validation verification, causality,

workflow: edge computing, compression, online learning, federated learning, augmented intelligence,

AI learns the world is, not the way it should be. (bias)
AI leans the world from the data presented, not the way it is.

AI algorithm may needs FDA styled drug trails and approval.

AI/ML microscopy in materials: predicting crystal structure by merging data mining and quantum physics. Chemical space is non-differentiable. Chemical space is a graph. Functionalitties at the nodes are defined within the context, such biological context, water, etc.  In many cases, chemical properties are hard to predict. Materials design can be thought as a search problem. Use mean field descriptors. Build precision micrscopy to map atoms?! Open data. Jupyter papers.
Localization: CNN, precision: Gaussian
theory-experiment matching.
Hypothesis driven science = forward mode P(dat/theory) P(theory)based on domain exptersize,
Q: how to get training data at atomic levels?


AI in health, Gina Tourassi
 Johnson, KW, J Am Colle Cardiol, 2018, 7 23,
https://www.sciencedirect.com/science/article/pii/S0735109718344085
AI in memogram scan by MIT/MGH
basal carcinoma, by deep learning,

genes and biology are responsbble for 10% of our health and well being.

modeling health instead of disease.

current opintin in biotechnology, 2019, v58, by Eberhart Voit
https://www.sciencedirect.com/science/article/pii/S0958166918301915





















No comments:

Post a Comment