Cloudy with a Chance of AI is back! And in this webinar our AI residents, Andrew van Aken and Michael McCarthy, will update you on everything Deep Learning as they discuss its application and latest updates, including practical examples of Ogilvy working in this space.

1. Powered by
What’s Next:
Cloudy with a
Chance of AI

2. Welcome
José Arteaga
Creative Digital Strategist
Ogilvy Consulting
Andrew Van Aken
Senior Consultant
Ogilvy
Michael McCarthy
Senior Consultant
Ogilvy

3. Tell us
where you
are dialing
in from!
What’s the weather
like in your city?

4. Do you
want this
deck?
It will be available for download
shortly after the webinar on:
slideshare.net/socialogilvy
Ogilvy staff: It’s also on
The Market!
themarket.ogilvy.com

5. Deep Learning

6. Source: Deep Learning for Python
Deep Learning

7. Nothing New
• Originally proposed in the 1940s by Warren McCulloch and Walter
Pitts
• Ideas refined in “Perceptrons”, published in 1969

10. ImageNet - 2012

11. Source: AIINDEX.org
• “These attendance
numbers show that
research focus has
shifted from symbolic
reasoning to machine
learning and deep
learning”
On the Rise

12. Deep
Learning - In
Practice

13. • A foundational model for deep learning is artificial neural networks
(ANN’s).
• Neural networks use “nodes”, which are interconnected and mimic the
neurons in a human brain.
Neural Networks

14. Neural Networks
• A foundational model for deep learning is artificial neural networks
(ANN’s).
• Neural networks use “nodes”, which are interconnected and mimic the
neurons in a human brain.
Nodes
Input Hidden Output

15. Neural Networks
Nodes
Input Hidden Output
Independent
Variables
Features
Dependent
Variable
• A foundational model for deep learning is artificial neural networks
(ANN’s).
• Neural networks use “nodes”, which are interconnected and mimic the
neurons in a human brain.

16. Neural Networks
• Nodes are interconnected and feed “forward” to an output

17. Neural Networks
Input Layer Output Layer
Hidden Layer
• Nodes are interconnected and feed “forward” to an output

18. Neural Networks
• Neural networks are models that mimic the way the human brain
interprets patterns - through signal processing and feature extraction.
Input Layer Output Layer
Hidden Layer
Signal
Feature Extraction
Output

19. How does it work?
Signal
Feature Extraction
Output

20. How does it work?
Signal
Feature Extraction
Output
Wheel

21. How does it work?
Signal
Feature Extraction
Output
Wheel

22. How does it work?
Signal
Feature Extraction
Output
Size

23. How does it work - Deep Learning
• Deep learning adds an additional layer of complexity versus standard
neural network models, allowing for more feature extraction.
Signal

24. How does it work - Deep Learning
• Deep learning adds an additional layer of complexity versus standard
neural network models, allowing for more feature extraction.
Signal
Wheel

25. How does it work - Deep Learning
• Deep learning adds an additional layer of complexity versus standard
neural network models, allowing for more feature extraction.
Signal
Feature Extraction
SizeWheel

26. How does it work - Deep Learning
• Deep learning adds an additional layer of complexity versus standard
neural network models, allowing for more feature extraction.
Signal
Feature Extraction
Output
SizeWheel Silhouette

27. How does it work - Deep Learning
Feature Extraction
Wheel Silhouette
…
…
…
• The “depth” in deep learning, refers to the many layers that are required
for advanced AI applications.

28. Training the Model
• In order to learn “features”, some deep learning models use
backpropagation.
Signal
Feature Extraction
Output

29. Training the Model
• In order to learn “features”, some deep learning models use
backpropagation.
Signal
Feature Extraction
Output

30. Training the Model
• In order to learn “features”, some deep learning models use
backpropagation.
Signal
Feature Extraction
Output

31. Training the Model
• In order to learn “features”, some deep learning models use
backpropagation.
Signal
Feature Extraction
Output

32. Training the Model
• In order to learn “features”, some deep learning models use
backpropagation.
Signal
Feature Extraction
Output

33. Summary
• Neural networks mimic the human brain
• Deep Learning models are neural networks with many, many layers
• Image recognition models extract features
• Deep Learning models are trained through backpropagation

34. Buzzword?

35. “Behind the surge is Google’s
growing investment in artificial
intelligence, particularly “deep
learning,” a technique whose
ability to make sense of images
and other data is enhancing
services like search and
translation” -MIT Tech Review
Buzzword?

36. • Image/Voice Recognition
• Text Generation
• Alexa/Watson/You Name It
• Almost everywhere…
Use Cases

37. Self Driving Cars

38. VIOLA:
Why, Salisbury must find his flesh and thought
That which I am not aps, not a man and in fire,
To show the reining of the raven and the wars
To grace my hand reproach within, and not a fair are hand,
That Caesar and my goodly father's world;
When I was heaven of presence and our fleets,
Text Generation

39. Art Generation

40. • “Using deep learning
algorithms trained on data
from 284,335 patients, we
were able to predict CV
risk factors from retinal
images with surprisingly
high accuracy”
Healthcare

42. What Does it Mean for Marketers?

43. Real Time Bidding

44. Logo Detection

45. Ad Tagging

46. New Methods

47. Future Trends

48. Accesible

49. Amazon Rekognition is based on the same proven,
highly scalable, deep learning technology
Accesible

50. Accesible

51. Mobile

52. Hardware

53. We don’t know…

54. Case Study

55. Webby

56. • “Melanoma kills 1 Australian every 6 hours. It is our deadliest skin
cancer but, ironically, one of the easiest to treat… survival rates can
top 98%. The key is identifying melanoma early.”
Melanoma

57. • IBM published a 2017 paper demonstrating state of the art accuracy
when applying deep learning to melanoma tumor detection
• Can get to 85% accuracy when using high scale images
Watson Health

58. Watson

59. The Question
• Would Australians actually engage with AI?

61. • Generated 4.88 million engagements from 45.6 million media
impressions
KPIs

62. • 22% of people who met Watson had an on-site skin check
• 40% were then referred onto a dermatologist for further analysis
• Some had suspicious moles identified then removed, like Ian Kelsall
KPIs

63. Going Forward
• Researchers to use the data to hopefully get to 90%+ accuracy

64. Link to Work
• https://www.ogilvy.com.au/our-work/outthink-melanoma

65. Questions?
José Arteaga
Creative Digital Strategist
Ogilvy Consulting
Andrew Van Aken
Senior Consultant
Ogilvy
Michael McCarthy
Senior Consultant
Ogilvy

66. Thank you.