Skip to main content

CS7641-Introduction

ML is the ROX Playlist

Prove theorem

What's ML?

Charles: Broader notion of building computational artifacts that learn over time based on experiences.

The math, science, engineering, computing behind it.

Michael: Computational applied statistics: Data, analyze the data, glean something from data via computational structures.

Induction and deduction

  • Induction applies example to more general rules. (ML depends on it.)
  • Deduction applies general rules into specific examples.

Approximate function induction.

Three major topics

Supervised Learning (approximation)

Taking labelled data set, glean information from it so that you can glean new data set.

Functional approximation: assume fundamental function to explain the real world.

Taken a data set with labels + functions to generalize the functions beyond the data you've seen.

Induction: specifics to generic (inductive bias)

Deduction: rules to specifics.

Unsupervised Learning(concise description)

Input data set, derive the structure from them.

Pixels -> Description(UL) -> Summaries -> Function approximation(SL) -> labels.

  1. Equally good to use features as summaries.

Reinforcement Learning

Generic AI to build agents.

RL is Learning from the delayed rewards. Playing game without knowing the rules.

Summary

Optimization

SL: labels data well

UL: cluster scores well

RL: behavior scores well

Algorithm central versus Data central (co-equal?)