DeepMind深度学习及前沿进展

1 . Deep Learning The AI revolution and its frontiers Nando de Freitas, Matt Hoffman, Ziyu Wang, Brendan Shillingford, Caglar Gulcehre, Misha Denil, Scott Reed, Serkan Cabi, Tobias Pfaff, Tom Paine, Yannis Assael, Yusuf Aytar, Yutian Chen,, Sergio Gomez, David Budden, Natasha Jaques 

2 .Hard coding has its limits 

3 . The machine learning way Faces Not faces 

4 . The machine learning way Faces Not faces 

5 . ? MNIST = Mixed National Institute of Standards and Technology - Download the dataset at http://yann.lecun.com/exdb/mnist/ 

6 .Hubel & Wiesel 

7 .A simple neural net for classifying images of digits 784 pixels 28x28 pixels ... Z softmax ... 0 1 2 9 Computed probabilities Y 0.1 0.0 0.0 0.2 0.0 0.1 0.1 0.0 0.4 0.1 0 1 2 3 4 5 6 7 8 9 [Modified @martin_gorner slide by from Martin Gorner] 

8 .The modular (layer-wise) view of neural nets 

9 . Supervised learning 0 1 2 3 4 5 6 7 8 9 0 0 0 0 0 0 0 0 1 0 Actual probabilities, “one-hot” encoded Cross entropy: Predicted probabilities Y 0.1 0.0 0.0 0.2 0.0 0.1 0.1 0.0 0.4 0.1 0 1 2 3 4 5 6 7 8 9 [Modified @martin_gorner slide by from Martin Gorner] 

10 . Supervised learning 0 1 2 3 4 5 6 7 8 9 0 0 0 0 0 0 0 0 1 0 Actual probabilities, “one-hot” encoded Cross entropy: Test loss Predicted probabilities Y 0.1 0.0 0.0 0.2 0.0 0.1 0.1 0.0 0.4 0.1 Training loss 0 1 2 3 4 5 6 7 8 9 [Modified @martin_gorner slide by from Martin Gorner] 

11 . Want better results? Go bigger and train end-2-end ! 784 200 Layer of neurons 100 60 30 10 softmax 0 1 2 ... 9 @martin_gorner 

12 . Case study: Visual speech recognition ● Improving speech recognition in noisy environments. ● Silent interfaces. ● Helping patients with aphonia. ● Resolving multi-talker simultaneous speech. 

13 .Step 1: Create a massive labelled dataset 

14 .Step 2: Train a large deep network 

15 .Step 3: Evaluate performance Ground truth: I am making this video for all my friends and family who may not know what I'm going through having done their research or even people that are actually going through what I'm going through Prediction: and making this video for all my friends and family you may not know what I'm going to do a little research or even people that actually going through what I'm going through Word Error Rate: 22.86% 

16 .Thousands of applications [Stanford] 

17 .The key AI revolution ingredients 1. Science 2. Data 3. Computing 4. Software 

18 .Deep learning research is like playing with lego Not like this But rather like this Combinatorial re-use is robust and amazing for creativity General Artificial Intelligence 

19 .Modularity backprop 

20 .Modularity backprop 

21 .Modularity backprop 

22 .Example: Linear layer 

23 .Another view as to why Deep Learning is so successful: Deep nets represent algorithms learned from data Differentiable Forth Interpreter Sebastian Riedel et al Jacob Andreas Resnets 

24 .Neural-Programmer Interpreters (NPI) 

25 .Frontiers that excite me 1. Reinforcement learning 2. Meta learning 3. Imitation 4. Robotics 5. Concepts and abstraction 6. Awareness and consciousness 7. Causal reasoning 

26 .Reinforcement Learning Framework OBSERVATIONS GOAL Agent Environment ACTIONS 

27 .Atari with deep RL 

28 .

29 .Alphazero