Deep Learning. Yee Whye Teh (Oxford Statistics & DeepMind)

Transcription

1 Deep Learning Yee Whye Teh (Oxford Statistics & DeepMind)

2 What is Machine Learning? Information Structure Prediction Decisions Actions data

3 What is Machine Learning? Information Structure Prediction Decisions Actions data

4 Learning Parameterised Functions 1 = arg min n nx i=1 L(y i,f (x i )) + k k

5 Learning Parameterised Functions 1 = arg min n nx i=1 L(y i,f (x i )) + k k! Modern deep learning frameworks allow for construction and learning of parameterised functions.! Consists of basic building blocks composed into computation graphs.! Highly expressive and flexible.! Modular: reusable complex building blocks are themselves composed of simpler building blocks.

6 Learning Parameterised Functions 1 = arg min n nx i=1 L(y i,f (x i )) + k k! Modern deep learning frameworks allow for construction and learning of parameterised functions.! Consists of basic building blocks composed into computation graphs.! Highly expressive and flexible.! Modular: reusable complex building blocks are themselves composed of simpler building blocks.! Computation graph structure expresses prior knowledge.

7 Learning Parameterised Functions 1 = arg min n nx i=1 L(y i,f (x i )) + k k! Modern deep learning frameworks allow for construction and learning of parameterised functions.! Consists of basic building blocks composed into computation graphs.! Highly expressive and flexible.! Modular: reusable complex building blocks are themselves composed of simpler building blocks.! Computation graph structure expresses prior knowledge.! Learning using stochastic gradient descent (on multiple CPUs, GPUs, clusters) is automated.

8 Artificial Neural Networks y x h y = (W 2 h + b 2 ) h = (W 1 x + b 1 )

9 Building Blocks! Linear/fully-connected/dense! sigmoid x 7! Wx+ b (x) = exp( x)! softmax = softmax(x 1,...,x n ) exp(x1 ) P i exp(x i),..., P exp(x n ) i exp(x i)! Losses! tanh tanh(x) = exp(x) exp( x) exp(x)+exp( x) CrossEntropy(t, y) = X i Square(t, y) =kt yk 2 2 t i log y i! relu relu(x) = max(0,x) Hinge(t, y) = max(0, 1 t y)

10 Building Blocks! Convolution! max pooling

11 Convolutional Networks (Convnets)! Both filter banks and layers are 4D tensors.

12 Hierarchy of Parts

13 Visual Processing in the Brain

14 Sequence Models! Natural language processes! Genomics PRP VBZ NNS IN DET NN She sells seashells by the seashore

15 Recurrent Neural Networks

16 Recurrent Neural Networks

17 Recurrent Neural Networks

18 Recurrent Neural Networks h t = (W h z t + b h ) z t = tanh(w z z t 1 + W x x t + b z )

19 Long Short Term Memory (LSTM)

20 Machine Translation with seq2seq!

21 GoogLeNet Architecture

22 Image Caption Generation black, orange and white cat laying on some paper on a desk. cat with mussed up fur sitting discontentedly on a messy desk. a cat lazily sits in the middle of a cluttered desk. a cat sitting on top of a pile of papers on a desk. a dark multicolored cat laying on a table cluttered with various items.

23 Show Attend and Tell!

24 Show Attend and Tell

25 Gradient Descent! Patrick Rebeschini will introduce optimization for machine learning later in the afternoon.! Iterative procedure: (t+1) = (t) t 1 n nx i=1 rl(y i,f (t)(x i )) + rd( (t) k 0 )!! Two questions:! scalability to large data sets?! how to compute derivatives?

26 Stochastic Gradient Descent! Estimate gradient of loss using minibatches of data: (t+1) = (t) t 1 B t X i2b t rl(y i,f (t)(x i )) + rd( (t) k 0 )! Reduce computation cost from O(n) to O( B t ).! More data is always better, as long as you have the compute to handle it.! Stochastic gradients are unbiased estimates convergence theory.! Stochasticity can help regularise and alleviate over-fitting!

27 Automatic i 8intermediate nodes h j 8input nodes x j 8intermediate nodes h j 8output nodes f k

28 Automatic Differentiation! Two major approaches: forward mode, and reverse mode i 8intermediate nodes h j 8input nodes x j 8intermediate nodes h j 8output nodes f k

29 Automatic Differentiation! Two major approaches: forward mode, and reverse mode i 8intermediate nodes h j 8input nodes x j 8intermediate nodes h j 8output nodes f k! Forward: O(#inputs*#nodes). Reverse: O(#outputs*#nodes).

30 Infrastructure! Infrastructure support critical to deep learning (and ML in general):! software frameworks allow fast model building, automating away most low-level operations.! Culture of sharing code via open source releases.! hardware allows fast training, and scalable productionisation.! large datasets and difficult challenges pushing frontier forward.

31 VAE in Keras/TensorFlow Colab Demo

32 Deep Learning est mort. Vive Differentiable Programming! - Yann LeCun Yeah, Differentiable Programming is little more than a rebranding of the modern collection Deep Learning techniques, the same way Deep Learning was a rebranding of the modern incarnations of neural nets with more than two layers. The important point is that people are now building a new kind of software by assembling networks of parameterized functional blocks and by training them from examples using some form of gradient-based optimization.it s really very much like a regular program, except it s parameterized, automatically differentiated, and trainable/optimizable.

33 More Resources! Tutorials and courses:! compgi22_advanced_deep_learning_and_reinforcement_learning/! Summer schools: MLSS, DLSS, RLSS! Conferences: NIPS, ICML, UAI, AISTATS! Journals: JMLR! ArXiv