Incremental Learning Chris Mesterharm Fordham University.

19
Incremental Learning Chris Mesterharm Fordham University

Transcript of Incremental Learning Chris Mesterharm Fordham University.

Page 1: Incremental Learning Chris Mesterharm Fordham University.

Incremental Learning

Chris Mesterharm

Fordham University

Page 2: Incremental Learning Chris Mesterharm Fordham University.

Artificial Intellegence

• Make machines as smart or smarter than humans

• Intelligent machines can automate many tasks performed by humans

• Research started at beginning of modern computer era

Page 3: Incremental Learning Chris Mesterharm Fordham University.

Programming AI

• Difficult to program machines to be intelligent

• Some success with well structured problems

• Less success with “simpler” skills such as perception and vision

Page 4: Incremental Learning Chris Mesterharm Fordham University.

Machine Learning

• Teach instead of program machines

• Similar to humans and animals

But how do we model learning?

Page 5: Incremental Learning Chris Mesterharm Fordham University.

Induction

• The sun has risen every day in the past

• The sun will rise tomorrow

Learning from examples

Page 6: Incremental Learning Chris Mesterharm Fordham University.

Induction Examples

• Predict promoter sequences in DNA that specify where a gene is starting

• Predict the amount of change in the price of a particular stock in the market

• Predict whether or not an image contains a picture of a bicycle

• Predict tomorrow’s weather

Page 7: Incremental Learning Chris Mesterharm Fordham University.

Stock Market Example

• Goal is to predict how price of IBM stock will change tomorrow

• Learning algorithm looks for a prediction rule using any information that might be relevant– Previous prices and sales volume

– Economic indicators and index fund prices

– Analysts predictions and recent new reports

• All the information is collected together as an instance of the learning problem.

Page 8: Incremental Learning Chris Mesterharm Fordham University.

Stock Market Example

• Goal is to label a new instance with a price that is close to the future price

• We already have many labeled instances from previous days

• We use the previously labeled instances to learn a rule that works well on future instances

Page 9: Incremental Learning Chris Mesterharm Fordham University.

Batch LearningInstance label

Algorithm rule

ruleInstance label

Page 10: Incremental Learning Chris Mesterharm Fordham University.

Incremental Learning

• Assume you get feedback on your predictions

• Refine your rule with label feedback

• Allows the rule to adjust to changing conditions

Page 11: Incremental Learning Chris Mesterharm Fordham University.

Online Learning

• Algorithm has no instances at start

• Algorithm gets a single instance each trial

• A trial is composed of three steps– Get instance– Predict label– Discover true label and use that information to

refine the prediction rule

Page 12: Incremental Learning Chris Mesterharm Fordham University.

Online Learning

ruleInstance Predicted label

Algorithm True label

Page 13: Incremental Learning Chris Mesterharm Fordham University.

Online Learning Examples

• Financial market• Weather prediction• Sports• Predicting the future• Spam detection• Cooking

Page 14: Incremental Learning Chris Mesterharm Fordham University.

WinnowLearns a hyperplane to separate instances

x

x

x

x

xx

x

o

o

o

o

o

o

o

oo

o

x

x

x

Page 15: Incremental Learning Chris Mesterharm Fordham University.

Winnow

• Strong theoretical guarantees

• Performs well even with large instances

• Robust to noise• Can learn a changing

hyperplane

Page 16: Incremental Learning Chris Mesterharm Fordham University.

Acquisitions: Reuters

Page 17: Incremental Learning Chris Mesterharm Fordham University.

Tracking Winnow Example

Page 18: Incremental Learning Chris Mesterharm Fordham University.

Expanding and Testing

• Algorithm should allow– Delays in label feedback– New ways to use information in instances– Ability to internally generate labels

• Test online algorithms on– Financial problems– Robotics problems

Page 19: Incremental Learning Chris Mesterharm Fordham University.

Conclusion

• Online learning is a theoretically strong model for learning

• Online learning allows an incremental and adaptive style of learning that might be closer to the way humans and animals learn

• Winnow is a successful example of online learning that works well in practice