C3_W1.pdf

Introduction to ML Strategy

Untitled

If you choose poorly, it is entirely possible that you end up spending six months charging in some direction only to realize after six months that that didn't do any good

Orthogonalization

The most effective machine learning people is they're very clear-eyed about what to tune in order to try to achieve one effect. This is a process we call orthogonalization.

By having orthogonal controls that are ideally aligned with the things you actually want to control, it makes it much easier to tune the knobs you have to tune.

Untitled

We need one knob to tune each step of the performance steps

  1. Try bigger network or different optimization algorithm
  2. Try regularization, or a bigger training set
  3. Get bigger dev set
  4. Change either the dev set or the cost function

Setting up your Goal

Single Number Evaluation Metric

The new thing that i tried is working better or worse than my last idea?

Having a single real number evaluation metric allows you to quickly tell if classifier A or classifier B is better

Precision: True positives (Of the examples that the classifier recognize as cats, percentage that are actual cats)

Recall: Of all the images that are cats, percentage of actual cats are correctly recognized