Skip to main content

Machine Learning for Healthcare

The Gist​

  • The model is never separate from the task.
  • Data are generated, not given.
  • Representation is a scientific claim.
  • Generalization is contextual.
  • Modern AI does not remove classical baselines.
  • Deployment creates new distributions.

Thinking about these, meditating over them, realizing them in practice (esp. learning by making mistakes), is far more important than knowing the specifics of some ‘hot’ new model or approach.

What is the question you’re trying to answer? What are you trying to do?

Are your labels actually measuring what you think/say they’re measuring?

Do you understand the information you’ve gathered/presented to you to solve this problem? Do you understand its projection as data? How did you represent it? Do you understand what you’ve lost representing it as such?

TODO​

  • Poisson
  • AIC and BIC
  • Algos Page
  • ARIMA, SARIMA, and GARCH
  • Baby Loss
  • Categorical Distro
  • Chebyshev Inequality
  • Conjugate Prior
  • Cox Prop Hazards in depth
  • Data / Inference / Interaction —> Informatics Solution
  • Discuss leakage
  • Feynman — Information
  • Gaussian Mixture Model
  • Huber Loss
  • Joint Distro
  • Latent variable models for discovering states in data — ?
  • Normalization Term?
  • PREVALANCE DICTATES DEMAND
  • Prior Distro?
  • Regression - Why MSE and not MAE?
  • Average gradient across all samples = Average of Expected Value of some samples — linearity of expectation. This makes thigns computationally tractable. Explain!
  • Traditional versys bayesian linear regression
  • Transformers and Attention