Neural Networks. William Cohen [pilfered from: Ziv; Geoff Hinton; Yoshua Bengio; Yann LeCun; Hongkak Lee - NIPs 2010 tutorial ]

Transcription

1 Neural Networks William Cohen [pilfered from: Ziv; Geoff Hinton; Yoshua Bengio; Yann LeCun; Hongkak Lee - NIPs 2010 tutorial ]

2 WHAT ARE NEURAL NETWORKS?

3 William s notation Logis;c regression + 1 / if y 1+ exp( x i w) i =1 - - P(y i x i,w),% 1 ( 0 -' 1 * if y & 1+ exp( x i w) i = 0-.- ) 1-1 logistic(u) 1+ e u

4 Predict pos on (x1,x2) iff ax1 + bx2 + c >0 = S(ax1 + bx2 + c)

5 On beyond linear classifiers What about data that s not linearly separable?

6 One powerful extension to logis;c regression: mul;layer networks Classifier is a mul;layer network of logis-c units Each unit takes some inputs and produces one output using a logis;c classifier Output of one unit can be the input of another Input layer Hidden layer Output layer 1 w 0,2 w 0,1 w 1,1 v 1 =S(w T X) w 1 x 1 w 1,2 w 2,1 w 2 z 1 =S(w T V) x 2 w 2,2 v 2 =S(w T X)

7 Mul;layer Logic Regression: one type of Ar;ficial Neural Network (ANN) Neuron is a cell in the brain Highly connected to other neurons, and performs computa;ons by combining signals from other neurons Outputs of these computa;ons may be transmieed to one or more other neurons

8 Does AI need ANNs? Neurons take ~ second to change state We can do complex scene recogni;on tasks in ~ 0.1 sec or about 100 steps of computa;on ANNs are similar: lots of parallelism not many steps of computa;on

9 BACKPROP: LEARNING FOR NEURAL NETWORKS

10 Ideas from Linear and Logis;c regression Define a loss which is squared error But over a network of logis;c units Minimize loss with gradient descent J X,y (w) = i ( y i y ˆ i ) 2 But output is network output Quick review: the math for linear regression and logis;c regression

11 Gradient Descent for Linear Regression (simplified) Differentiate the loss function: J X,y (w) = 1 2 J(w) = 1 w j w j 2 i ( y i y ˆ i 2 1 ) = 2 = y i ŷ i i i ( ) = y i ŷ i i ( ) = y i ŷ i i ( ) ( y i ŷ i ) i w j w j x j $ & y i % 2 ŷ i j w j x j j predict with : ˆ y i = ' w j x j ) ( 2 n j w j x j

12 Gradient Descent for Logis;c Regression (simplified) Output of a logistic unit: After some math:

13 ( f n )'= nf n 1 f ' p 13

14 Gradient Descent for Logis;c Regression (simplified) Output of a logistic unit: After some math: 1 S(g) = 1+ e g S g = g(1 g)

15 DERIVATION OF BACKPROP

16 Nota;on a y becomes a target t a is an ac;va;on and net is a weighted sum of inputs i.e., a=s(net) S is sigmoid func;on w kj is a weight from hidden j to output k w ji is a weight from input i to hidden j J(w) = 1 2 w ji ( t k a ) 2 k k w kj

17 Deriva;on: gradient for output- layer weights J(w) = 1 2 k J = J w kj a k ( t k a ) 2 k a k net k net k w kj

18 Deriva;on: output layer J(w) = 1 2 k J = J w kj a k net k w kj = w kj ( t k a ) 2 k a k net k net k w kj j' w kj' a j' = a j J a k = ( t k a ) k a k net = a ( 1 a ) k k From logistic regression result

19 Deriva;on: output layer J(w) = 1 2 k ( t k a ) k J = ( t k a ) k a ( k 1 a ) k a j w kj 2 δ k ( t k a k )a ( k 1 a ) k error w kj J = δ k a j sensitivity

20 Deriva;on: gradient for hidden- layer weights J(w) = 1 2 J(w) = 1 2 J(w) = 1 2 w ji J = # % $ k k k k ( t k a ) k # # t k S % % $ $ 2 j w kj a j && (( '' # # # &&& t k S % w kj S% w ji a i ( % ( $ $ j $ i ' ( '' J a k a k net k net k a j 2 2 & a j net j ( ' net j w ji

21 Deriva;on: hidden layer J = w ji " $ # k k J a k a k net k net k a j " % = $ δ k w kj 'a j (1 a j )a i # & δ j ( δ k w ) kj a j 1 a j k J w ji = δ j a i % a j net j ' & net j w ji ( )

22 Compu;ng the weight update For nodes k in output layer: δ k ( t k a ) k a ( k 1 a ) k For nodes j in hidden layer: δ j ( δ k w ) kj a j 1 a j k For all weights: ( ) w kj = w kj ε δ k a j Propagate errors backward BACKPROP Can carry this recursion out further if you have multiple hidden layers w ji = w ji ε δ j a i

23 Some extensions to BackProp. Different units: Can replace (1 + e - x ) - 1 with tanh, Different architectures: Can extend this method to any directed graph Weight sharing/parameter ;e - ing Can have the same weights appear in mul;ple loca;ons in the graph (more later)

24 EXPRESSIVENESS OF ANNS

25 Networks of logis;c units can do a lot One logis;c unit can implement and AND or an OR of a subset of inputs e.g., (x 3 AND x 5 AND AND x 19 ) Every boolean func;on can be expressed as an OR of ANDs e.g., (x 3 AND x 5 ) OR (x 7 AND x 19 ) OR So one hidden layer can express any BF (But it might need lots and lots of hidden units)

26 Networks of logis;c units can do a lot An ANN with one hidden layer can also approximate every bounded con;nuous func;on. (But it might need lots and lots of hidden units)

27 Limita;ons of Backprop Learning is slow: thousands of itera;ons Learning is oden ineffec;ve on deep networks more than 1-2 hidden layers (excep;on: convolu;onal nets) Lots of parameters to tweak structure of the network regulariza;on, learning rate, momentum terms, number of itera;ons/convergence, common trick is early stopping: train ;ll error on a held- out valida;on set stops decreasing similar to L2- regulariza;on BP finds a local minima, not a global one star;ng point maeers (mul;ple restarts) warning: you usually want to randomize the weights when you start It only is useful for supervised learning tasks: or, is it?

28 DEEP NEURAL NETWORKS

29 Observa;ons and mo;va;ons Slow is not scary anymore Moore s law, GPUs, Brains seem to be hierarchical deeper more abstract features computed from simpler ones 100 steps!= 3 steps

30 CONVOLUTIONAL NETS

31 weight- sharing and convolu;onal networks Convolu;onal neural networks for vision tasks: These can be trained with Backprop, even with many hidden layers Every unit is a feature detector for part of the re;na, with a receptor field that defines its inputs and (conceptually) a loca-on on the re;na Two types of units: detec;on: learn to recognize some feature (copied to mul;ple places via weight- sharing) pooling: combine detectors from mul;ple nearby loca;ons together (e.g., max, sampling, )

32 Model of vision in animals

33 Vision with ANNs

34 Vision with ANNs

35 Similar technique applies to audio

36 DENSITY MODELLING WITH ANNS

37 Neural network auto- encoding Assume we would like to learn the following (trivial?) output func;on: Using the following network: Input Output Can this be done?

38 Note that each value is assigned to the edge from the corresponding input Learned parameters

39 Hypothe;cal example (not actually from an autoencoder) Data Reconstruction Data Reconstruction

40 Neural network autoencoding The hidden layer is a compressed version of the data Reconstruc-on error on a vector x is related to P(x) on the probability that the auto- encoder was trained with. Denoising auto- encoders: trained to reconstruct x from a noisy version of x

41 Density Modeling with ANNs Sigmoid belief nets (Neal, 1992): explicitly model a generative process. Top layer nodes are generated by a binomial. Sampling from (generating data with) a SBN: Sample h k from P(h k ) Sample from h k-1 from P(h k-1 h k ) Sample x from P(x h 1) What about learning?

42 Recall gradient for logis;c regression This can be interpreted as a difference between the expected value of y x j =1 in the data and the expected value of y x j =1 as predicted by the model Gradient ascent tries to make those equal So: If we can sample from the model we can approximate the gradient. = E x,y~data [y x] E x~data,y~ ˆP(y x) [y x] 42

43 Density Modeling with ANNs Sigmoid belief nets (Neal, 1992): explicitly model a generative process. Top layer nodes are generated by a binomial. Sampling from (generating data with) a SBN: Sample h k from P(h k ) Sample from h k-1 from P(h k-1 h k ) Sample x from P(x h 1)

44 Density Modeling with ANNs Notation: x is now visible units v Restricted Boltzmann machine: equivalent to a very deep sigmoid belief network, with lots of weights tied. Usually show this as a two-layer network with one hidden layer and symmetric weights. RBNs can compute Pr(v h) or Pr(h v) directly

45 Compu;ng probabili;es with RBM via Gibbs sampling j j j j 0 < v i hj> < v i h j > a fantasy i i i i t = 0 t = 1 t = 2 t = infinity Start with any value on the visible units. Then alternate between updating the hidden units (in parallel) and the visible units in parallel. System will converge to a stream of <v,h> drawn from Pr(v,h) Goal of learning: push system s Pr(v,h) toward the observed probabilities Start with observed x, let system drift away, then train it to stay closer to the observed x. variant - clamp some units to a fixed value c, and sample from Pr(v 1,h v 2 =c)

46 Training a stack of RBNs =

47 Training a stack of RBNs = Step 1: train an RBN Step 2:

48 Training a stack of RBNs a deep Boltzmann machine = Steps 3, : repeat as needed. Then fine-tune with BP on a discriminative task: clamp n-1 of the inputs to x and use Gibbs to sample from Pr(y x)

49 DBN examples: learning the digit 2 16 x 16 = 256 pixel image; 50 x 256 hidden units

50 Example: reconstruc;ng digits with a network trained on different copies of 2 Data Reconstruction from activated binary features Data Reconstruction from activated binary features New test images from the digit class that the model was trained on Images from an unfamiliar digit class (the network tries to see every image as a 2)

51 Samples generated a_er clamping the label y

52 PSD features learned from all digits [LeCunn]

53 DENSITY MODELLING COMBINED WITH BACKPROP

54 Training a stack of RBNs a deep Boltzmann machine = Steps 3, : repeat as needed. Then fine-tune with BP on a discriminative task:

55 Deep learning methods Greedily learn a deep model of the unlabeled data. Deep Boltzmann machines, denoising autoencoders, PSDs, Learn a hidden layer from the inputs; Fix this layer and learn a deeper layer from it; Then fine-tune with BP. Another approach for RBNs: unroll the stack of RBNs into a multi-layer network, add appropriate output notes, and use it starting point for BP.

56 0 1 9.

57

58 Error reduced from 16% to 6% by using the deep features.

59 Modeling documents using top 2000 words.

60

61

62 What you should know What a neural network is and what sorts of things it can compute Why it s not linear Backprop: loss func;on, updates, etc. Limita;ons of Backprop without pre- training on deep networks