5. Performances
2012: from 26% to 16% error rate in image
classification with convolutional neural nets Similar disruption observed in speech
6. What’s the secret sauce of Deep
Neural Nets ?
● RMSProp
● ADAM
● SGD
● Sigmoid
● Tanh
● ReLU
● Leaky ReLU
● CNN
● RNN
● LSTM
● GAN
● WaveNet
● PathNet
The more,
the cleaner
and well
distributed
the better
7. How does it works?
From details captured in the first layers to higher level of abstraction in deeper layers
10. Input = “A small yellow bird with a black crown and a short black pointed beak”
Which images are real, which images are generated by the algorithm?
From classification to generation
11. A small yellow bird with a back crown and a short black pointed beak
From classification to generation
Answer: None of these images ever existed. They are all generated by the algorithm.
StackGAN, Han Zhang & al, paper here
12. GAN
Generative Adversarial Networks
Ian J. Goodfellow & al, 2014
StackGAN
Text to Photo-realistic Image Synthesis
Han Zhang & al, 2016
From classification to generation
14. Unsurpervised learning
Reinforcement Learning
→ predict a scalar reward
→ A few bits per samples
Supervised Learning = on-sight navigation
→ Predicting human supplied data
→ 10 to 10000 bits per samples
All the rest is unsupervised Learning = offshore navigation
→ Predicting unknown parts from observation
→ Building its own representation
→ Millions of bits per samples
Ex: predicting frames in videos, generating images, music...
Any Limits?