Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.
Session
Ganes Kesari
Gramener
How Deep Learning is Saving the Planet
Session
2
INTRODUCTION
Co-founder & Head
of Analytics
“Simplify Data
Science for all”
100+ Clients
Insights as Stories
Help apply ...
3
OUR STORY BEGINS 50,000 YEARS AGO…
What happened to them?
https://wattsupwiththat.com/2017/01/20/humans-not-climate-chan...
4
Humans are ecological serial killers…“
http://www.wrbh.org/wp-content/uploads/2018/05/Spaiens-book-cover.jpg
…even with ...
5
CAN WE SAVE OUR BIODIVERSITY?
6
SPOTTING, IDENTIFYING AND COUNTING ANIMALS TO SAVE THEM
Gramener has partnered with Microsoft AI for Earth
https://www.m...
7
MACHINE LEARNING 101
New Input
Desired
Outcome
Machine learning
how to do the job
Known Input
Known
Outcome
“Programs th...
8
WHY DEEP LEARNING?
8
Input Output
Identify features
to teach model
Traditional Machine Learning
Deep Learning
Person
Nam...
9
NOT VERY DIFFERENT FROM HOW WE LEARN
Training ..Versatile Detection!
10
WHERE HAVE THE SALMON GONE?
Source: Giphy (https://media.giphy.com/media/QM5GJO6J8lDfa/giphy.mp4)
11
CAN YOU IDENTIFY THESE SALMON SPECIES?
Sockeye Steelhead
12
MONITORING SALMON MOVEMENT
13
BUILDING THE MODEL – FASTER RCNN
https://tryolabs.com/blog/2018/01/18/faster-r-cnn-down-the-rabbit-hole-of-modern-objec...
14
CLASSIFYING THE 12 SALMON SPECIES
15
THE MODEL IN ACTION
Video
16
A VISUAL NARRATIVE OF THE ENGAGEMENT
Microsoft published case study on this project at: https://partner.microsoft.com/e...
17
SPOTTING ELEPHANTS IN THE WILD
Source: Giphy (https://media.giphy.com/media/1AHZzdXVTYDtJVTO5a/giphy.mp4)
18
CAN YOU SPOT THE ELEPHANTS?
https://www.savetheelephants.org/project/tsavo-aerial-defence/
19
THE DATA & CHALLENGES
Annotate
20
MODEL BUILDING
Region Proposal Network Region of Interest Pooling Regional CNN
FASTER RCNN
SSD MULTIBOX
21
MODEL RESULTS
22
CAN WE CLASSIFY ALL SPECIES ON EARTH?
Source: Giphy (https://media.giphy.com/media/CFk1wEH7Cke0E/giphy.mp4)
23
INATURALIST: FOSTERING CITIZEN SCIENCE
https://play.google.com/store/apps/details?id=org.inaturalist.android&hl=en
24
BUILDING THE MODEL
Inception v4
25
LIVE API DEMO
26
LET’S SAVE THE PENGUINS
Penguin populations are at risk
in Antarctica and researchers
need help to detect how it’s
redu...
27
TRACKING THE PENGUINS
https://www.zooniverse.org/projects/penguintom79/penguin-watch/about/research
~100 cameras in 16 ...
28
CROWD-SOURCED ANNOTATIONS
https://www.zooniverse.org/projects/penguintom79/penguin-watch/classify
29
APPROACHES TO COUNTING CROWDS
Occlusion
Density Difference
Perspective Distortion
Camera angle
30
COUNTING USING DENSITY-BASED ESTIMATIONS
Preserve spatial information
Localize count
Handle scale variations
No longer ...
31
APPROACH IN BRIEF
https://arxiv.org/pdf/1707.09605.pdf
Input an image Density Map Estimate the countSplit into 9 patches
32
CHALLENGES WITH THE DATA
Cleaned dataset:
• Training: 18k
• Validation: 3k
• Test: 9k
Hurdles
• Camera angles,
• Occlus...
33
MODEL ARCHITECTURE
• High-level prior to classify image into buckets
• Density estimation to create the density map
• N...
34
WALKTHROUGH
35
DEEP LEARNING TAKEAWAYS: WHEN THE RUBBER HITS THE ROAD
• Acquire & clean data
• Label your own data
• Look out for prac...
36
The first step towards change is awareness. The
second is acceptance.
– Nathaniel Branden
“
WildMe projects - https://w...
37
Session
@kesaritweetsgramener.com @kesari
Presentation deck with references at
gkesari.com/intelligentcloud
Event
partners
Expo
partners
Expo light
partners
Upcoming SlideShare
Loading in …5
×

2019 Intelligent Cloud Conf - How Deep Learning is Saving the Planet

121 views

Published on

Presented by Ganes Kesari at the Intelligent Cloud Conference, Copenhagen on 9-Apr-2019 (http://intelligentcloud.dk)

Published in: Data & Analytics
  • Be the first to comment

2019 Intelligent Cloud Conf - How Deep Learning is Saving the Planet

  1. 1. Session Ganes Kesari Gramener How Deep Learning is Saving the Planet Session
  2. 2. 2 INTRODUCTION Co-founder & Head of Analytics “Simplify Data Science for all” 100+ Clients Insights as Stories Help apply & adopt Analytics Our data science platform, Gramex is now open- sourced!
  3. 3. 3 OUR STORY BEGINS 50,000 YEARS AGO… What happened to them? https://wattsupwiththat.com/2017/01/20/humans-not-climate-change-wiped-out-australian-megafauna/
  4. 4. 4 Humans are ecological serial killers…“ http://www.wrbh.org/wp-content/uploads/2018/05/Spaiens-book-cover.jpg …even with stone-age tools, our ancestors wiped out half the planet’s large terrestrial mammals. - Yuval Noah Harari
  5. 5. 5 CAN WE SAVE OUR BIODIVERSITY?
  6. 6. 6 SPOTTING, IDENTIFYING AND COUNTING ANIMALS TO SAVE THEM Gramener has partnered with Microsoft AI for Earth https://www.microsoft.com/en-us/ai/ai-for-earth
  7. 7. 7 MACHINE LEARNING 101 New Input Desired Outcome Machine learning how to do the job Known Input Known Outcome “Programs that solve the problem” “Programs that learn to solve the problem” vs
  8. 8. 8 WHY DEEP LEARNING? 8 Input Output Identify features to teach model Traditional Machine Learning Deep Learning Person Name Input Output Model automatically identifies features to learn Person Name https://www.cs.toronto.edu/~ranzato/publications/taigman_cvpr14.pdf
  9. 9. 9 NOT VERY DIFFERENT FROM HOW WE LEARN Training ..Versatile Detection!
  10. 10. 10 WHERE HAVE THE SALMON GONE? Source: Giphy (https://media.giphy.com/media/QM5GJO6J8lDfa/giphy.mp4)
  11. 11. 11 CAN YOU IDENTIFY THESE SALMON SPECIES? Sockeye Steelhead
  12. 12. 12 MONITORING SALMON MOVEMENT
  13. 13. 13 BUILDING THE MODEL – FASTER RCNN https://tryolabs.com/blog/2018/01/18/faster-r-cnn-down-the-rabbit-hole-of-modern-object-detection/ Region Proposal Network Region of Interest Pooling Regional CNN
  14. 14. 14 CLASSIFYING THE 12 SALMON SPECIES
  15. 15. 15 THE MODEL IN ACTION Video
  16. 16. 16 A VISUAL NARRATIVE OF THE ENGAGEMENT Microsoft published case study on this project at: https://partner.microsoft.com/en-us/case-studies/gramener/
  17. 17. 17 SPOTTING ELEPHANTS IN THE WILD Source: Giphy (https://media.giphy.com/media/1AHZzdXVTYDtJVTO5a/giphy.mp4)
  18. 18. 18 CAN YOU SPOT THE ELEPHANTS? https://www.savetheelephants.org/project/tsavo-aerial-defence/
  19. 19. 19 THE DATA & CHALLENGES Annotate
  20. 20. 20 MODEL BUILDING Region Proposal Network Region of Interest Pooling Regional CNN FASTER RCNN SSD MULTIBOX
  21. 21. 21 MODEL RESULTS
  22. 22. 22 CAN WE CLASSIFY ALL SPECIES ON EARTH? Source: Giphy (https://media.giphy.com/media/CFk1wEH7Cke0E/giphy.mp4)
  23. 23. 23 INATURALIST: FOSTERING CITIZEN SCIENCE https://play.google.com/store/apps/details?id=org.inaturalist.android&hl=en
  24. 24. 24 BUILDING THE MODEL Inception v4
  25. 25. 25 LIVE API DEMO
  26. 26. 26 LET’S SAVE THE PENGUINS Penguin populations are at risk in Antarctica and researchers need help to detect how it’s reducing over the years Source: Giphy (https://giphy.com/gifs/push-bwLowbhUWm2lO)
  27. 27. 27 TRACKING THE PENGUINS https://www.zooniverse.org/projects/penguintom79/penguin-watch/about/research ~100 cameras in 16 locations, and hourly images over years Crowd-sourced annotations to identify the penguins
  28. 28. 28 CROWD-SOURCED ANNOTATIONS https://www.zooniverse.org/projects/penguintom79/penguin-watch/classify
  29. 29. 29 APPROACHES TO COUNTING CROWDS Occlusion Density Difference Perspective Distortion Camera angle
  30. 30. 30 COUNTING USING DENSITY-BASED ESTIMATIONS Preserve spatial information Localize count Handle scale variations No longer looking for a head!
  31. 31. 31 APPROACH IN BRIEF https://arxiv.org/pdf/1707.09605.pdf Input an image Density Map Estimate the countSplit into 9 patches
  32. 32. 32 CHALLENGES WITH THE DATA Cleaned dataset: • Training: 18k • Validation: 3k • Test: 9k Hurdles • Camera angles, • Occlusion, • Perspective distortion, • Density difference, • Weather conditions Work done by Gramener in partnership with Microsoft AI for Earth
  33. 33. 33 MODEL ARCHITECTURE • High-level prior to classify image into buckets • Density estimation to create the density map • NC6 v3 virtual machine with V100 GPU card • Trained for 200 epochs, MAE for the model : ~10.5 https://arxiv.org/pdf/1707.09605.pdf Work done by Gramener in partnership with Microsoft AI for Earth
  34. 34. 34 WALKTHROUGH
  35. 35. 35 DEEP LEARNING TAKEAWAYS: WHEN THE RUBBER HITS THE ROAD • Acquire & clean data • Label your own data • Look out for practical data challenges • Don’t stop at counts - go for actionability • Build it into the user’s natural workflow • Sensitize users on low accuracy • Plan for model refresh
  36. 36. 36 The first step towards change is awareness. The second is acceptance. – Nathaniel Branden “ WildMe projects - https://www.wildbook.org/ https://www.flukebook.org/ THE WAY FORWARD..
  37. 37. 37 Session @kesaritweetsgramener.com @kesari Presentation deck with references at gkesari.com/intelligentcloud
  38. 38. Event partners Expo partners Expo light partners

×