C3W1 ML Strategy

Question 1

Key ideas how to improve performance

Answer 1

Collect more data
Diverse training set
Train longer
Try ADAM
Try bigger network
Try smaller network
Try dropout
Try L2 regularisation
Change network architecture (activation function, hidden units)

Question 2

Precision and recall

Answer 2

Precision - (of all recognitions as cats, how many actual cats?)

Recall - what percent of all cats are correctly recognised?

Question 3

F1 score

Answer 3

~Average of precision and recall
Harmonic mean
2 / (1/p + 1/r)

Question 4

Optimising and satisfying metrics

Answer 4

Optimising - accuracy - the lower the better
Satisfying - running time (set threshold- less than specific amount)
You should have only one optimising metric

Question 5

Two steps

Answer 5

1. Define metric
2. How to do well on the metric

Question 6

Base optimal error

Answer 6

The best level of performance

Question 7

How to achieve human level performance?

Answer 7

Get labeled data from humans
Gain insight from manual error analysis
Better analys of bias variance