Advice for applying Machine Learning
Andrew Ng Stanford University
Andrew Y. Ng
Today’s Lecture
• Advice on how getting learning algorithms to different applications.
• Most of today’s material is not very mathematical. But it’s also some of the hardest material in this class to understand.
• Some of what I’ll say today is debatable.
• Some of what I’ll say is not good advice for doing novel machine learning research.
• Key ideas:
1. Diagnostics for debugging learning algorithms.
2. Error analyses and ablative analysis.
3. How to get started on a machine learning problem.
– Premature (statistical) optimization.
Andrew Y. Ng
Debugging Learning Algorithms
Andrew Y. Ng
Motivating example:
Debugging learning algorithms
• Anti-spam. You carefully choose a small set of 100 words to use as features. (Instead of using all 50000+ words in English.)
• Bayesian logistic regression, implemented with gradient descent, gets 20% test error, which is unacceptably high.
• What to do next?
Andrew Y. Ng
• Bayesian logistic regression:
Fixing the learning algorithm
• Common approach: Try improving the algorithm in different ways. – Try getting more training examples.
– Try a smaller set of features.
– Try a larger set of features.
– Try changing the features: Email header vs. email body features. – Run gradient descent for more iterations.
– Try Newton’s method.
– Use a different value for λ.
– Try using an SVM.
• This approach might work, but it’s very time-consuming, and largely a matter of luck whether you end up fixing what the problem really is.
Andrew Y. Ng
Diagnostic for bias vs. variance
Better approach:
– Run diagnostics to figure out what the problem is. – Fix whatever the problem is.
Bayesian logistic regression’s test error is 20% (unacceptably high).
Suppose you suspect the problem is either:
– Overfitting (high variance).
– Too few features to classify spam (high bias).
Diagnostic:
– Variance: Training error will be much lower than test error. – Bias: Training error will also be high.
Andrew Y. Ng
error
Typical learning curve for high variance:
More on bias vs. variance
m (training set size)
• Test error still decreasing as m increases. Suggests larger training set will help.
• Large gap between training and test error.
Test error
Training error
Desired performance
Andrew Y. Ng
error
Typical learning curve for high bias:
More on bias vs. variance
• Even training error is unacceptably high.
• Small gap between training and test error.
m (training set size)
Test error
Training error
Desired performance
Andrew Y. Ng
Diagnostics tell you what to try next
Bayesian logistic regression, implemented with gradient descent.
Fixes to try:
– Try getting more training examples.
– Try a smaller set of features.
– Try a larger set of features.
– Try email header features.
– Run gradient descent for more iterations. – Try Newton’s method.
– Use a different value for λ.
– Try using an SVM.
Fixes high variance. Fixes high variance. Fixes high bias. Fixes high bias.
Andrew Y. Ng
Optimization algorithm diagnostics
• Bias vs. variance is one common diagnostic.
• For other problems, it’s usually up to your own ingenuity to construct your own diagnostics to figure out what’s wrong.
• Another example:
– Bayesian logistic regression gets 2% error on spam, and 2% error on non-spam. (Unacceptably high error on non-spam.)
– SVM using a linear kernel gets 10% error on spam, and 0.01% error on non- spam. (Acceptable performance.)
– But you want to use logistic regression, because of computational efficiency, etc. • What to do next?
Andrew Y. Ng
Objective
More diagnostics
• Other common questions:
– Is the algorithm (gradient descent for logistic regression) converging?
Iterations
It’s often very hard to tell if an algorithm has converged yet by looking at the objective.
J(θ)
Andrew Y. Ng
(weights w(i) higher for non-spam than for spam). – Bayesian logistic regression? Correct value for λ?
– SVM? Correct value for C?
More diagnostics
• Other common questions:
– Is the algorithm (gradient descent for logistic regression) converging? – Are you optimizing the right function?
– I.e., what you care about:
Andrew Y. Ng
Let θSVM be the parameters learned by an SVM.
Let θBLR be the parameters learned by Bayesian logistic regression. You care about weighted accuracy:
θSVM outperforms θBLR. So: BLR tries to maximize: Diagnostic:
Diagnostic
An SVM outperforms Bayesian logistic regression, but you really want to deploy Bayesian logistic regression for your application.
Andrew Y. Ng
Case 1:
Case 2:
Two cases
But BLR was trying to maximize J(θ). This means that θBLR fails to maximize J, and the problem is with the convergence of the algorithm. Problem is with optimization algorithm.
This means that BLR succeeded at maximizing J(θ). But the SVM, which does worse on J(θ), actually does better on weighted accuracy a(θ).
This means that J(θ) is the wrong function to be maximizing, if you care about a(θ). Problem is with objective function of the maximization problem.
Andrew Y. Ng
Diagnostics tell you what to try next
Bayesian logistic regression, implemented with gradient descent.
Fixes to try:
– Try getting more training examples.
– Try a smaller set of features.
– Try a larger set of features.
– Try email header features.
– Run gradient descent for more iterations. – Try Newton’s method.
– Use a different value for λ.
– Try using an SVM.
Fixes high variance.
Fixes high variance.
Fixes high bias.
Fixes high bias.
Fixes optimization algorithm. Fixes optimization algorithm. Fixes optimization objective. Fixes optimization objective.
Andrew Y. Ng
The Stanford Autonomous Helicopter
Payload: 14 pounds Weight: 32 pounds
Andrew Y. Ng
1. Buildasimulatorofhelicopter.
Machine learning algorithm
2. Choose a cost function. Say J(θ) = ||x – xdesired||2 (x = helicopter position)
3. Runreinforcementlearning(RL)algorithmtoflyhelicopterinsimulation,so as to try to minimize cost function:
θRL = arg minθ J(θ)
Suppose you do this, and the resulting controller parameters θRL gives much worse
performance than your human pilot. What to do next?
Improve simulator? Modify cost function J? Modify RL algorithm?
Andrew Y. Ng
Simulator
The controller given by θRL performs poorly. Suppose that:
Debugging an RL algorithm
1. Thehelicoptersimulatorisaccurate.
2. TheRLalgorithmcorrectlycontrolsthehelicopter(insimulation)soasto
minimize J(θ).
3. MinimizingJ(θ)correspondstocorrectautonomousflight.
Then: The learned parameters θRL should fly well on the actual helicopter. Diagnostics:
1. If θRL flies well in simulation, but not in real life, then the problem is in the simulator. Otherwise:
2. Let θhuman be the human control policy. If J(θhuman) < J(θRL), then the problem is in the reinforcement learning algorithm. (Failing to minimize the cost function J.)
3. IfJ(θhuman)≥J(θRL),thentheproblemisinthecostfunction.(Maximizingit doesn’t correspond to good autonomous flight.)
Andrew Y. Ng
More on diagnostics
• Quite often, you’ll need to come up with your own diagnostics to figure out what’s happening in an algorithm.
• Even if a learning algorithm is working well, you might also run diagnostics to make sure you understand what’s going on. This is useful for:
– Understanding your application problem: If you’re working on one important ML application for months/years, it’s very valuable for you personally to get a intuitive understand of what works and what doesn’t work in your problem.
– Writing research papers: Diagnostics and error analysis help convey insight about the problem, and justify your research claims.
– I.e., Rather than saying “Here’s an algorithm that works,” it’s more interesting to say “Here’s an algorithm that works because of component X, and here’s my justification.”
• Good machine learning practice: Error analysis. Try to understand what your sources of error are.
Andrew Y. Ng
Error Analysis
Andrew Y. Ng
Face detection
Eyes segmentation Nose segmentation
Logistic regression
Label
Error analysis
Many applications combine many different learning components into a “pipeline.” E.g., Face recognition from images: [contrived example]
Camera image
Preprocess (remove background)
Mouth segmentation
Andrew Y. Ng
Camera image
Preprocess (remove background)
Face detection
Eyes segmentation Nose segmentation
Logistic regression
Label
How much error is attributable to each of the components?
Overall system
85%
Plug in ground-truth for each component, and see how accuracy changes.
Face detection Eyes segmentation Nose segmentation Mouth segmentation Logistic regression
91% 95% 96% 97%
Conclusion: Most room for improvement in face detection and eyes segmentation.
Error analysis
Mouth segmentation
Component
Accuracy
Preprocess (remove background)
85.1%
And1re0w0Y%. Ng
Ablative analysis
Error analysis tries to explain the difference between current performance and perfect performance.
Ablative analysis tries to explain the difference between some baseline (much poorer) performance and current performance.
E.g., Suppose that you’ve build a good anti-spam classifier by adding lots of clever features to logistic regression:
– Spelling correction.
– Sender host features.
– Email header features.
– Email text parser features.
– Javascript parser.
– Features from embedded images.
Question: How much did each of these components really help?
Andrew Y. Ng
Ablative analysis
Simple logistic regression without any clever features get 94% performance.
Just what accounts for your improvement from 94 to 99.9%?
Ablative analysis: Remove components from your system one at a time, to see how it breaks.
Component Accuracy
Overall system 99.9% Spelling correction 99.0
Sender host features 98.9% Email header features 98.9%
Email text parser features 95% Javascript parser 94.5% Features from images 94.0%
[baseline]
Conclusion: The email text parser features account for most of the improvement.
Andrew Y. Ng
Getting started on a learning problem
Andrew Y. Ng
Approach #1: Careful design.
Getting started on a problem
• Spend a long term designing exactly the right features, collecting the right dataset, and designing the right algorithmic architecture.
• Implement it and hope it works.
• Benefit: Nicer, perhaps more scalable algorithms. May come up with new, elegant, learning algorithms; contribute to basic research in machine learning.
Approach #2: Build-and-fix.
• Implement something quick-and-dirty.
• Run error analyses and diagnostics to see what’s wrong with it, and fix its errors.
• Benefit: Will often get your application problem working more quickly. Faster time to market.
Andrew Y. Ng
Camera image
Preprocess (remove background)
This system’s much too complicated for a first attempt.
Premature statistical optimization
Very often, it’s not clear what parts of a system are easy or difficult to build, and which parts you need to spend lots of time focusing on. E.g.,
Face detection
Logistic regression
Label
Eyes segmentation Nose segmentation
Step 1 of designing a learning system: Plot the data.
Mouth segmentation
The only way to find out what needs work is to implement something quickly, and find out what parts break.
[But this may be bad advice if your goal is to come up with new machine learning algorithms.]
Andrew Y. Ng
Mail delivery robot
Navigation
Obstacle avoidance
Complexity of non-Riemannian geometries
Robot manipulation
...
Convergence bounds for sampled non- monotonic logic
[Based on Papadimitriou, 1995]
Andrew Y. Ng
The danger of over-theorizing
Color invariance
Differential geometry of 3d manifolds
Object detection
VC dimension
3d similarity learning
Summary
Andrew Y. Ng
Summary
• Time spent coming up with diagnostics for learning algorithms is time well- spent.
• It’s often up to your own ingenuity to come up with right diagnostics.
• Error analyses and ablative analyses also give insight into the problem.
• Two approaches to applying learning algorithms: – Design very carefully, then implement.
• Risk of premature (statistical) optimization.
– Build a quick-and-dirty prototype, diagnose, and fix.
Andrew Y. Ng