Day 7 - PCA, Kernels, Ensembles

Posted on June 6, 2017 by Govind Gopakumar

Lecture slides in pdf form

Prelude

Announcements

Recap

Clustering

Generative modelling

Dimensionality Reduction

Aren’t more features better? - I

How many “features” should we have?

How useful are these features?

Aren’t more features better? - II

Curse of dimensionality

Feature extraction

Principal Components Analysis - I

Model overivew

What is an informative direction?

Principal Components Analysis - II

Review of Covariance

Geometry of covariances

Principal Components Analysis - III

Geometry of model

Computing the spread

Principal Components Analysis - IV

Solving it analytically

As a linear embedding

Principal Components Analysis - V

What is our loss function / optimization?

Where is the optima for this?

Principal Components Analysis - VI

Steps in PCA

How many to choose?

Principal Components Analysis - VII

Why is this useful?

Usage

Kernels

Increasing dimensionality of data - I

Wait, what?

Okay, how?

Increasing dimensionality of data - II

Procedure

Computational issues?

Increasing dimensionality of data - III

Using a Kernel

Examples of kernels

Increasing dimensionality of data - IV

Examples of kernels

Why are they useful?

Increasing dimensionality of data - V

How do we use them in models?

Examples

Boosting and ensembles

Ensemble models - I

What?

How?

Ensemble models - II

Bagging

Why would this work?

Ensemble models - III

Boosting

Process

Ensemble models - IV

AdaBoost

Can it learn complicated shapes?

Ensemble models - V

Comments

Why should we use either?

Conclusion

Concluding Remarks

Takeaways

Announcements

References