notes.txt

TODO list for next time (29 Oct 2014)

- Likelihood lecture: change cost function to "Absolute Deviation" rather than "Least Absolute Deviation"

- Bayesian Lecture: Motivate MCMC (perhaps plot 2D posterior first, but then talk about curse of dimensionality)

- MCMC motivation: add a graphical description of the process, and perhaps an animation

- Model Validation: 04.3 does not tell a compelling story! This section needs a lot of work. Ideas:

----------------------------------------------------------------

Model Validation:

Motivating Cross-validation -- digits

- Start w/ K Neighbors with no test set: "perfect" classification
- Next do a single train/test split... better, but we lose some data
- Introduce idea of holdout/K-fold cross-validation
- Show cross_val_score and GridSearch

Unbalanced Data: accuracy is not what you want!

- Show an unbalanced dataset
- Show a 95% accurate estimator... y_pred = 0
- Introduce precision (1 - contamination) and recall (completeness)
- Show how these can be used within cross_val_score & GridSearch

Cross-val works on regression problems too!

- Show on ??


Bias-Variance Tradeoff & Learning Curves

- Move current material to a new notebook.


----------------------------------------------


PCA lecture...

- Need to do a better job of linking together eigenvectors, principal components, dimensionality reduction, etc. How to do this?