22. Hot Dogs or Meatballs?
Image recognition is one of the most important
developments in AI. Close your eyes and imagine:
- Taking a shower
- Eating your lunch
- Driving
36. Be Aware of Image Hacking 1
Image Source: https://codewords.recurse.com/issues/five/why-do-neural-networks-think-a-panda-is-a-vulture
37. Be Aware of Image Hacking 2
Image Source: https://codewords.recurse.com/issues/five/why-do-neural-networks-think-a-panda-is-a-vulture
38. Be Aware of Image Hacking 3
Image Source: https://medium.com/@ageitgey/machine-learning-is-fun-part-8-how-to-intentionally-trick-neural-networks-b55da32b7196
39. Be Aware of Image Hacking - Fake Fakes
This could be a devastating example if the autonomous vehicle is trained
to drive over newspaper!
Pedestrian (98.5%) Newspaper (95.9%)
40. What can we do?
I need better training!
Any ideas?
41. Generate Training Data Idea 1
Augment source image by:
Rotation Perspective Transformation
50. Time Saving Tip
Generate many variations per source image
but:
● Don’t save to disk (slow and takes space)
● Run it under a Python Generator Object
○ Generate images as you train (fast and scalable)
51. Train a Winning Model
… and become a billionaire!
Image Source: HBO
A cast of 30 something techies continuously try to get funding for their ideas. They go to tech competitions, conferences but were always circumvented by their competitor, nemesis and billionaire CEO, Gavin Belson.silicon valley opening scene season 1
A cast of 30 something techies continuously try to get funding for their ideas. They go to tech competitions, conferences but were always circumvented by their competitor, nemesis and billionaire CEO, Gavin Belson.silicon valley opening scene season 1
A cast of 30 something techies continuously try to get funding for their ideas. They go to tech competitions, conferences but were always circumvented by their competitor, nemesis and billionaire CEO, Gavin Belson.silicon valley opening scene season 1
Erlich, the serial entrepreneur, runs an incubator from his house including this group of startup guys.
One of them is Jing Yang, the freeloader of the group.
Finally, Erlich through some miscommunication sold Jing Yang’s idea for a food identification app.
During the demo, the app only identifies hot dogs and all other foods as not hot dogs.
Perplexed to the success of the app. Dinesh wonders ...
Erlich had the idea to use the student from his friend’s stanford class to scrape the internet for food pictures to train Jing Yang’s model.
But the students steal his idea, rendering his company worthless.
Erlich again had his butt handed to him because he doesn’t have enough understanding how to train a neural network.
Erlich had the idea to use the student from his friend’s stanford class to scrape the internet for food pictures to train Jing Yang’s model.
But the students steal his idea, rendering his company worthless.
Erlich again had his butt handed to him because he doesn’t have enough understanding how to train a neural network.
See these perfect specimens of hot dog pictures taken probably by a professional food photographer.
For any neural network, training data is very important. See on a simple task of identifying a hot dog. How one simple hot dog can be represented?
See these perfect specimens of hot dog pictures taken probably by a professional food photographer.
For any neural network, training data is very important. See on a simple task of identifying a hot dog. How one simple hot dog can be represented?
For most activities, us, humans only need to use our eyes
- to drive, to eat, to catch a ball.
We haven’t needed radars, lidars, other electronic sensors.
I’ll show you some signs and you all recognize these. They’re shown under perfect conditions.
Life is not like that. Even for just the School Zone Sign. There are lots of them.
Some are very localized to the region. Please use shotguns only because a handgun would just be too dangerous!
Life is not like that. Even for just the School Zone Sign. There are lots of them.
Some are very localized to the region. Please use shotguns only because a handgun would just be too dangerous!
Life is not like that. Even for just the School Zone Sign. There are lots of them.
Some are very localized to the region. Please use shotguns only because a handgun would just be too dangerous!
Another way to get a robust classification neural net model is to use Dropout.
It sounds kind of crazy and un-intuitive. Why would you want to wipe out information in the middle of training your network?
Because you don’t want to model to count on any specific features because at anytime that feature may not be there.
Go forth, generate a gazillion training images for your winning model and become a billionaire!
Go forth, generate a gazillion training images for your winning model and become a billionaire!