- Discriminitive vs generative models
- Introduce popular datasets we will work with: Fisher irises, Old Faithful, NIST, faces
- Add more algorithmic details: computational complexity, etc.
- Make a table where minimization functions for all of the topics are summarized.
- Add conclusions at the end of EVERY lecture.
- Make demo style consistent (consider using SNS).
- Example with non-balanced dataset (terrorist/non-terrorist) is ambiguous, redo
- Too much attention to simple types, need more objects/classes
- add
np.einsum
consideration - think on some visualization for np.arrays
- error in image with slicing
- add list/array memory representation image
- fix cycle/ufunction time comparison (add comparison with list)
- add Polya urn models (Beta distribution etc)
- Problem with 2 boxes, bead, and pearl. 2 boxes, one contains 1 bead and the other 1 pearl. We don't know where pearl is (50/50). One more pearl is aced into box "B", shuffled, randomply pulled back. Pearl is pulled back. What is the probability to have pearl in "A"? in "B"?
- Add Logistic regression
- Mention orthogonal distance regression (scipy.odr)
- Regression (to mean) visualization
- Workbook, problem with error bars, add slider to change order of the fitting polynomial
- Workbook, add problem to demonstrate LAD robustness (compared to OLS)
- Set feaures on click in demo
- Coordinate descent (has connections to SVM and E-M algo)
- Use UI from GMM demo to set the initial point
- Use maximum descent algo to get 2D maximum in a matrix of integers as is done in https://www.youtube.com/watch?v=HtSuA80QTyo&t=37s
- Pull out not only fitting but classification function as well (from visualization).
- Add visualization as in http://setosa.io/ev/principal-component-analysis/
- Use idea with finding dissimilar image from https://dou.ua/lenta/articles/computer-vision-magic/
- Visualize surface (do Delaune triangulation and draw as a bunch of triangles)