Advice for Getting Models Work Mahdi Roozbahani Lecturer, - PowerPoint PPT Presentation

CX4242: Advice for Getting Models Work Mahdi Roozbahani Lecturer, Computational Science and Engineering, Georgia Tech These slides are adopted from Polo, Andrew w. Moore, and Vivek Srikumar, and Chao Zhang

Outline • Model Diagnostics • Error Analysis • Practical Advice 2

Debugging Machine Learning 3

Different Ways to Improve Your Model 4

First Step: Diagnose Your Model 5

Overfitting v.s. Underfitting 6

Overfitting (High Variance) 7

Underfitting (High Bias) 8

Different Ways to Improve Your Model 9

Diagnostics 10

Have Your Model Converged? 11

Have Your Model Converged? 12

Have Your Model Converged? 13

Different Ways to Improve Your Model 14

Diagnostics 15

What to Measure 16

Outline • Model Diagnostics • Error Analysis • Practical Advice 17

Machine Learning in Context 18

Error Analysis 19

Example: Text Processing System 20

Example: Text Processing System 21

Error Analysis 22

Outline • Model Diagnostics • Error Analysis • Practical Advice 23

Advice for ML Workflow in Practice 24

Advice for ML Workflow in Practice 25

What to Watch Out For? 26

What to Watch Out For? 27

What to Watch Out For? 28