learned representations

38
Learned Representations @ejlbell

Upload: eddie-bell

Post on 15-Apr-2017

316 views

Category:

Software


0 download

TRANSCRIPT

Page 1: Learned Representations

Learned Representations

@ejlbell

Page 2: Learned Representations

Me

2

Page 3: Learned Representations

Lyst

3

Page 4: Learned Representations

4

Feature engineering

Page 5: Learned Representations

5

Type of transform Examples

Unary exp, log, sqrt

Normalisation mean, variance

Aggregation count, sum, mean

Dimensional reduction PCA, clustering, manifold

Text tagging, parsing, stemming

Image histograms, key points, super pixels, segmentation

Others temporal / spatial

Page 6: Learned Representations

AI must fundamentally understand the

world around us and this can only be

achieved if it can learn to identify and

disentangle the underlying explanatory

factors hidden in the observed milieu of

low-level sensory data.

2014 - Representation Learning: A Review and New Perspectives. - Bengio et al.

Page 7: Learned Representations

Human ingenuity and prior knowledge

Feature Engineering

Representation Learning

Sufficiently powerful models that learn “good” feature transforms

Page 8: Learned Representations

8

Deep Learning

Page 9: Learned Representations

Image Filters

-1

-2

-1

0

0

0

+1

+2

+1

-1

0

+1

-2

0

+2

-1

0

+1

Page 10: Learned Representations

VGG

Representations

2015 - Very Deep Convolutional Networks for

Large-Scale Image Recognition. Simonyan and

Zisserman

Page 11: Learned Representations

Regularization

11

Page 12: Learned Representations

12

Representations

Page 13: Learned Representations

13

Cat

Dog

Male

Female

Page 14: Learned Representations

14

Cat

Dog

Male

Female

Page 15: Learned Representations

15

Cat

Dog

Male

Female

Page 16: Learned Representations

Content

Page 17: Learned Representations

Similar to ‘dress’

Page 18: Learned Representations
Page 19: Learned Representations
Page 20: Learned Representations

20

Applications

Page 21: Learned Representations
Page 22: Learned Representations

22

a group of young girls standing next

to each other on the beachA clock tower with a clock on top of it

Page 23: Learned Representations

A bunch of bananas hanging from a tree

Page 24: Learned Representations

24

Page 25: Learned Representations

25

Page 26: Learned Representations

26

+ =

Page 27: Learned Representations

27

Page 28: Learned Representations
Page 29: Learned Representations

29

Page 30: Learned Representations

30

Page 31: Learned Representations

31

Page 32: Learned Representations

32

Page 33: Learned Representations

33

Page 34: Learned Representations

34

Page 35: Learned Representations

35

1 0 0

1 0 0

0 1 0

0.7 -0.6 -8

0 0 1

0 0 1

0 1 0

-8 -0.6 0.7

Page 36: Learned Representations

But … not a magic bullet

36

* Expensive in terms of hardware

* Expensive in terms of time

* Expensive in terms of expertise

* Expensive in terms of labelled data

* Blackbox, can’t do inference

Page 37: Learned Representations

Thanks especially to all the people I stole this content from

Page 38: Learned Representations

Questions?

@ejlbell