Computer Science Department

Learning on the Fly:Rapid Adaptation to the Image

Erik Learned-Millerwith Vidit Jain, Gary Huang,

Laura Sevilla Lara, Manju Narayana, Ben Mears

2Learning on the Fly

“Traditional” machine learning

Learning happens from large data sets

• With labels: supervised learning

• Without labels: unsupervised learning

• Mixed labels: semi-supervised learning,transfer learning,learning from one (labeled) example,self-taught learning,domain adaptation

3Learning on the Fly

Learning on the Fly

Given:

• A learning machine trained with traditional methods

• a single test image (no labels)

Learn from the test image!

4Learning on the Fly

Learning on the Fly

Given:

• A learning machine trained with traditional methods

• a single test image (no labels)

Learn from the test image!

• Domain adaptation where the “domain” is the new image

• No covariate shift assumption.

• No new labels

5Learning on the Fly

An Example in Computer Vision

Parsing Images of Architectural ScenesBerg, Grabler, and Malik ICCV 2007.

• Detect easy or “canonical” stuff.

• Use easily detected stuff to bootstrap models of harder stuff.

6Learning on the Fly

Claim

This is so easy and routine for humans that it’s hard to realize we’re doing it.

• Another example…

7Learning on the Fly

Learning on the fly…

8Learning on the Fly

Learning on the fly…

9Learning on the Fly

Learning on the fly…

10Learning on the Fly

What about traditional methods…

Hidden Markov Model for text recognition:

• Appearance model for characters

• Language model for labels

• Use Viterbi to do joint inference

11Learning on the Fly

What about traditional methods…

Hidden Markov Model for text recognition:

• Appearance model for characters

• Language model for labels

• Use Viterbi to do joint inference

DOESN’T WORK!

Prob( |Label=A) cannot be well estimated, fouling up the whole process.

12Learning on the Fly

Lessons

We must assess when our models are broken, and use other methods to proceed….

• Current methods of inference assume probabilities are correct!

• “In vision, probabilities are often junk.”

• Related to similarity becoming meaningless beyond a certain distance.

13Learning on the Fly

2 Examples

Face detection (CVPR 2011)

OCR (CVPR 2010)

14Learning on the Fly

Preview of results: Finding false negatives

Viola-Jones Learning on the Fly

15Learning on the Fly

Eliminating false positives

Viola-Jones Learning on the Fly

16Learning on the Fly

Eliminating false positives

Viola-Jones Learning on the Fly

17Learning on the Fly

Run a pre-existing detector...

18Learning on the Fly

Run a pre-existing detector...

Key

Face

Non-face

Close to

boundary

19Learning on the Fly

Gaussian Process Regression

negative positive

learn smooth mapping

from appearance to score

apply mapping to borderline

patches

20Learning on the Fly

Major Performance Gains

21Learning on the Fly

Comments

No need to retrain original detector

• It wouldn’t change anyway!

No need to access original training data

Still runs in real-time

GP regression is done for every new image.

22Learning on the Fly

Noisy Document

We fine herefore t

linearly rolatcd to the

when this is calculated

equilibriurn. In short,

on the null-hypothesis:

Initial Transcription

23Learning on the Fly

Premise

We would like to fine confident wordsto build a document-specific model,but it is difficult to estimate Prob(error).

However, we can bound Prob(error).

Now, select words with

• Prob(error)<epsilon.

24Learning on the Fly

“Clean Sets”

25Learning on the Fly

Document specific OCR

Extract clean sets (error bounded sets)

Build document-specific models from clean set characters

Reclassify other characters in document

• 30% error reduction on 56 documents.

26Learning on the Fly

Summary

Many applications of learning on the fly.

Adaptation and bootstrapping new models is more common in human learning than is generally believed.

Starting to answer the question: “How can we do domain adaptation from a single image?”