[Structuring Machine Learning Projects] week1. ML Strategy (1)

[TOC]

I-Introduction to ML Strategy

Why ML Strategy

A lot of ideas of improving ML performance: strategy on how to choose.

→ how to figure out which ones to pursue and which ones to discard ?

Orthogonalization

How to tune hyperparams & what to expect.

TV tuning example: each knob does only one thing.

Chain of assumptions in ML:
training set performance → dev set → test set → real world

  • "one knob for each chain"
  • Will go through these "knobs" in this course.
  • Don't use early stopping: this is not orthogonalized enough

II-Setting up your goal

Single number evaluation metric

Faster progress if only have one single real number evaluation metric. → more efficient in making decisions.

example:
Using both percision and recall as metric is not good → difficule to pick the best model to keep on iterating from.
→ Use F1 score instead.

Satisficing and Optimizing metric

When it's difficult to pick a single real number eval metric → set up satisficing and optimizing metrics.

example: accuracy & running time trade-off
Instead of doing a linear combination of the two, use this:
maximize accuracy
subject to running time <= 100 ms
In this case: accuracy is optimizing metric, running time is satisficing metric.



In general:
if having N metrics → pick 1 as optimizing metric, the N-1 rest as satisficing metric.

example: assistant wake-up word accuracy VS false positive.

Train-dev-test distributions

How to setup dev/test sets.
Idea: dev/test sets come from the same distribution.

example: cat classification in different regions.
This is bad dev/test setup:

Good practice: random shuffle data and split into dev/test sets.


Size of the dev and test sets

Pre-DL era, old way of splitting data: 70/30 or 60/30/10 split.
→ resonable when datasets are small. (100~10k examples)

In DL era: much more training (~1M) examples.
⇒ 98/1/1 split is more resonable.

  • Size of test/dev set: big enough to give high confidence in system's performance.
  • OK to not have a test set, but not recommended.

When to change dev/test sets and metrics

example 1
cat classification: algo A has pornographic false positives.

→ change metric to penalize heavily pornographic FPs.

to implement this weighting, need to go through dev/test sets.




example 2
Cat classification: user's upload is blury while trained on high quality images.


III-Comparing to human-level performance

Why human-level performance?

Workflow ML can be more efficient when trying to match human level performance.

Bayes optional error: best possible error, theoritical optimal.


ML progress usually slows down after surpassing human-level performance:

  • usually human-level is not far from Bayes optimal
  • as long as ML performace < human, there are tools to improve performance.

Avoidable bias

Using human rating can prevent overfitting on training set.
exapmle: compare training set error with human performance.

  • err_train > err_human ⇒ focus on reducing bias (e.g. bigger NN)
  • err_train ~= err_human ⇒ focus on reducing variance (e.g. regularize, more training data)


→ Use human-level error as a proxy for Bayes error.

Terminology:

  • Avoidable bias is the gap between training err and Bayes err.
    (interpretataion: some errors are inavoidable because Bayes err is not 0.)

  • Variance: gap between training err and dev err.

Understanding human-level performance

"Human level error as proxy for Bayes error"

example: Medical image classification.

⇒ Should pick lowest human error as an estimate (upper bound) of Bayes error.

Error analysis example (which human-err to pick to estimate avoidable bias):

Surpassing human-level performance

What's the avoidable bias when err_train and err_dev are smaller than err_human ?
→ less clear in choosing directions.

examples of tasks where ML >> human performance:

⇒ all these tasks are:

  • learned from structured data
  • are not natural perception tasks
  • have processed huge amount of data

Improving your model performance

Recall: two fundamental assumptions of supervised learning:

  • You can fit training set well (achieve to ~= avoidable bias)
  • The performance on training set generalize well to dev/test sets. (achieve low variance)

The big roadmap:

comments powered by Disqus