Lecture 03: Machine Learning for Language Technology - Linear Classifiers
5 likes2,736 views
The document outlines an introductory machine learning course focused on linear classifiers relevant to language technology, detailing various models such as perceptrons and support vector machines (SVMs). It covers essential concepts including feature representation, training algorithms, and classification methods, highlighting the importance of maximizing the margin between classes in supervised learning. Key algorithms and their applications are summarized, including logistic regression and the margin-infused relaxed algorithm (MIRA).
![Preliminaries
Examples
x is a name, y is a label classifying the name
f0(x, y) =
8
<
:
1 if x contains “George”
and y = “Person”
0 otherwise
f1(x, y) =
8
<
:
1 if x contains “Washington”
and y = “Person”
0 otherwise
f2(x, y) =
8
<
:
1 if x contains “Bridge”
and y = “Person”
0 otherwise
f3(x, y) =
8
<
:
1 if x contains “General”
and y = “Person”
0 otherwise
f4(x, y) =
8
<
:
1 if x contains “George”
and y = “Object”
0 otherwise
f5(x, y) =
8
<
:
1 if x contains “Washington”
and y = “Object”
0 otherwise
f6(x, y) =
8
<
:
1 if x contains “Bridge”
and y = “Object”
0 otherwise
f7(x, y) =
8
<
:
1 if x contains “General”
and y = “Object”
0 otherwise
x=General George Washington, y=Person → f(x, y) = [1 1 0 1 0 0 0 0]
x=George Washington Bridge, y=Object → f(x, y) = [0 0 0 0 1 1 1 0]
x=George Washington George, y=Object → f(x, y) = [0 0 0 0 1 1 0 0]
Machine Learning for Language Technology 8(55)](https://image.slidesharecdn.com/lecture03ml4ltmarinasantini2013-130830021004-phpapp02/85/Lecture-03-Machine-Learning-for-Language-Technology-Linear-Classifiers-8-320.jpg)
![Preliminaries
Block Feature Vectors
x=General George Washington, y=Person → f(x, y) = [1 1 0 1 0 0 0 0]
x=George Washington Bridge, y=Object → f(x, y) = [0 0 0 0 1 1 1 0]
x=George Washington George, y=Object → f(x, y) = [0 0 0 0 1 1 0 0]
One equal-size block of the feature vector for each label
Input features duplicated in each block
Non-zero values allowed only in one block
Machine Learning for Language Technology 9(55)](https://image.slidesharecdn.com/lecture03ml4ltmarinasantini2013-130830021004-phpapp02/85/Lecture-03-Machine-Learning-for-Language-Technology-Linear-Classifiers-9-320.jpg)
![Linear Classifiers
Linear Classifiers - Bias Terms
Often linear classifiers presented as
y = arg max
y
m
j=0
wj × fj (x, y) + by
Where b is a bias or offset term
But this can be folded into f
x=General George Washington, y=Person → f(x, y) = [1 1 0 1 1 0 0 0 0 0]
x=General George Washington, y=Object → f(x, y) = [0 0 0 0 0 1 1 0 1 1]
f4(x, y) =
1 y =“Person”
0 otherwise
f9(x, y) =
1 y =“Object”
0 otherwise
w4 and w9 are now the bias terms for the labels
Machine Learning for Language Technology 11(55)](https://image.slidesharecdn.com/lecture03ml4ltmarinasantini2013-130830021004-phpapp02/85/Lecture-03-Machine-Learning-for-Language-Technology-Linear-Classifiers-11-320.jpg)
![Linear Classifiers
Perceptron
Choose a w that minimizes error
w = arg min
w t
1 − 1[yt = arg max
y
w · f(xt, y)]
1[p] =
1 p is true
0 otherwise
This is a 0-1 loss function
Aside: when minimizing error people tend to use hinge-loss or
other smoother loss functions
Machine Learning for Language Technology 17(55)](https://image.slidesharecdn.com/lecture03ml4ltmarinasantini2013-130830021004-phpapp02/85/Lecture-03-Machine-Learning-for-Language-Technology-Linear-Classifiers-17-320.jpg)
![Linear Classifiers
Aside: Min error versus max log-likelihood (i)
Highly related but not identical
Example: consider a training set T with 1001 points
1000 × (xi , y = 0) = [−1, 1, 0, 0] for i = 1 . . . 1000
1 × (x1001, y = 1) = [0, 0, 3, 1]
Now consider w = [−1, 0, 1, 0]
Error in this case is 0 – so w minimizes error
[−1, 0, 1, 0] · [−1, 1, 0, 0] = 1 > [−1, 0, 1, 0] · [0, 0, −1, 1] = −1
[−1, 0, 1, 0] · [0, 0, 3, 1] = 3 > [−1, 0, 1, 0] · [3, 1, 0, 0] = −3
However, log-likelihood = −126.9 (omit calculation)
Machine Learning for Language Technology 41(55)](https://image.slidesharecdn.com/lecture03ml4ltmarinasantini2013-130830021004-phpapp02/85/Lecture-03-Machine-Learning-for-Language-Technology-Linear-Classifiers-41-320.jpg)
![Linear Classifiers
Aside: Min error versus max log-likelihood (ii)
Highly related but not identical
Example: consider a training set T with 1001 points
1000 × (xi , y = 0) = [−1, 1, 0, 0] for i = 1 . . . 1000
1 × (x1001, y = 1) = [0, 0, 3, 1]
Now consider w = [−1, 7, 1, 0]
Error in this case is 1 – so w does not minimizes error
[−1, 7, 1, 0] · [−1, 1, 0, 0] = 8 > [−1, 7, 1, 0] · [0, 0, −1, 1] = −1
[−1, 7, 1, 0] · [0, 0, 3, 1] = 3 < [−1, 7, 1, 0] · [3, 1, 0, 0] = 4
However, log-likelihood = -1.4
Better log-likelihood and worse error
Machine Learning for Language Technology 42(55)](https://image.slidesharecdn.com/lecture03ml4ltmarinasantini2013-130830021004-phpapp02/85/Lecture-03-Machine-Learning-for-Language-Technology-Linear-Classifiers-42-320.jpg)
Lecture 03: Machine Learning for Language Technology - Linear Classifiers
- 1. Machine Learning for Language Technology Uppsala University Department of Linguistics and Philology Slides borrowed from previous courses. Thanks to Ryan McDonald (Google Research) and Prof. Joakim Nivre Machine Learning for Language Technology 1(55)
- 2. Introduction Linear Classifiers Classifiers covered so far: Decision trees Nearest neighbors Next two lectures: Linear classifiers Statistics from Google Scholar (Sept 2013): “Maximum Entropy”&“NLP”; 11,400 hits (2013), 2,660 (2009), 141 before 2000 “SVM”&“NLP”: 10,900 hits (2013), 2,210 (2009), 16 before 2000 “Perceptron”&“NLP”: 3,160 hits (2013), 947 (2009), 118 before 2000 All are linear classifiers that have become important tools in Language Technology in the past 10 years or so. Machine Learning for Language Technology 2(55)
- 3. Introduction Outline Today: Preliminaries: input/output, features, etc. Linear classifiers Perceptron Large-margin classifiers (SVMs, MIRA) Logistic regression Next time: Structured prediction with linear classifiers Structured perceptron Structured large-margin classifiers (SVMs, MIRA) Conditional random fields Case study: Dependency parsing Machine Learning for Language Technology 3(55)
- 4. Preliminaries Inputs and Outputs Input: x ∈ X e.g., document or sentence with some words x = w1 . . . wn, or a series of previous actions Output: y ∈ Y e.g., parse tree, document class, part-of-speech tags, word-sense Input/output pair: (x, y) ∈ X × Y e.g., a document x and its label y Sometimes x is explicit in y, e.g., a parse tree y will contain the sentence x Machine Learning for Language Technology 4(55)
- 5. Preliminaries Feature Representations We assume a mapping from input-output pairs (x, y) to a high dimensional feature vector f(x, y) : X × Y → Rm For some cases, i.e., binary classification Y = {−1, +1}, we can map only from the input to the feature space f(x) : X → Rm However, most problems in NLP require more than two classes, so we focus on the multi-class case For any vector v ∈ Rm, let vj be the jth value Machine Learning for Language Technology 5(55)
- 6. Preliminaries Features and Classes All features must be numerical Numerical features are represented directly as fi (x, y) ∈ R Binary (boolean) features are represented as fi (x, y) ∈ {0, 1} Multinomial (categorical) features must be binarized Instead of: fi (x, y) ∈ {v0, . . . , vp} We have: fi+0(x, y) ∈ {0, 1}, . . . , fi+p(x, y) ∈ {0, 1} Such that: fi+j (x, y) = 1 iff fi (x, y) = vj We need distinct features for distinct output classes Instead of: fi (x) (1 ≤ i ≤ m) We have: fi+0m(x, y), . . . , fi+Nm(x, y) for Y = {0, . . . , N} Such that: fi+jm(x, y) = fi (x) iff y = yj Machine Learning for Language Technology 6(55)
- 7. Preliminaries Examples x is a document and y is a label fj (x, y) = 1 if x contains the word“interest” and y =“financial” 0 otherwise fj (x, y) = % of words in x with punctuation and y =“scientific” x is a word and y is a part-of-speech tag fj (x, y) = 1 if x = “bank”and y = Verb 0 otherwise Machine Learning for Language Technology 7(55)
- 8. Preliminaries Examples x is a name, y is a label classifying the name f0(x, y) = 8 < : 1 if x contains “George” and y = “Person” 0 otherwise f1(x, y) = 8 < : 1 if x contains “Washington” and y = “Person” 0 otherwise f2(x, y) = 8 < : 1 if x contains “Bridge” and y = “Person” 0 otherwise f3(x, y) = 8 < : 1 if x contains “General” and y = “Person” 0 otherwise f4(x, y) = 8 < : 1 if x contains “George” and y = “Object” 0 otherwise f5(x, y) = 8 < : 1 if x contains “Washington” and y = “Object” 0 otherwise f6(x, y) = 8 < : 1 if x contains “Bridge” and y = “Object” 0 otherwise f7(x, y) = 8 < : 1 if x contains “General” and y = “Object” 0 otherwise x=General George Washington, y=Person → f(x, y) = [1 1 0 1 0 0 0 0] x=George Washington Bridge, y=Object → f(x, y) = [0 0 0 0 1 1 1 0] x=George Washington George, y=Object → f(x, y) = [0 0 0 0 1 1 0 0] Machine Learning for Language Technology 8(55)
- 9. Preliminaries Block Feature Vectors x=General George Washington, y=Person → f(x, y) = [1 1 0 1 0 0 0 0] x=George Washington Bridge, y=Object → f(x, y) = [0 0 0 0 1 1 1 0] x=George Washington George, y=Object → f(x, y) = [0 0 0 0 1 1 0 0] One equal-size block of the feature vector for each label Input features duplicated in each block Non-zero values allowed only in one block Machine Learning for Language Technology 9(55)
- 10. Linear Classifiers Linear Classifiers Linear classifier: score (or probability) of a particular classification is based on a linear combination of features and their weights Let w ∈ Rm be a high dimensional weight vector If we assume that w is known, then we define our classifier as Multiclass Classification: Y = {0, 1, . . . , N} y = arg max y w · f(x, y) = arg max y m j=0 wj × fj (x, y) Binary Classification just a special case of multiclass Machine Learning for Language Technology 10(55)
- 11. Linear Classifiers Linear Classifiers - Bias Terms Often linear classifiers presented as y = arg max y m j=0 wj × fj (x, y) + by Where b is a bias or offset term But this can be folded into f x=General George Washington, y=Person → f(x, y) = [1 1 0 1 1 0 0 0 0 0] x=General George Washington, y=Object → f(x, y) = [0 0 0 0 0 1 1 0 1 1] f4(x, y) = 1 y =“Person” 0 otherwise f9(x, y) = 1 y =“Object” 0 otherwise w4 and w9 are now the bias terms for the labels Machine Learning for Language Technology 11(55)
- 12. Linear Classifiers Binary Linear Classifier Divides all points: Machine Learning for Language Technology 12(55)
- 13. Linear Classifiers Multiclass Linear Classifier Defines regions of space: i.e., + are all points (x, y) where + = arg maxy w · f(x, y) Machine Learning for Language Technology 13(55)
- 14. Linear Classifiers Separability A set of points is separable, if there exists a w such that classification is perfect Separable Not Separable This can also be defined mathematically (and we will shortly) Machine Learning for Language Technology 14(55)
- 15. Linear Classifiers Supervised Learning – how to find w Input: training examples T = {(xt, yt)} |T | t=1 Input: feature representation f Output: w that maximizes/minimizes some important function on the training set minimize error (Perceptron, SVMs, Boosting) maximize likelihood of data (Logistic Regression) Assumption: The training data is separable Not necessary, just makes life easier There is a lot of good work in machine learning to tackle the non-separable case Machine Learning for Language Technology 15(55)
- 16. Linear Classifiers Perceptron ... The resulting hyperplane is called -perceptron- ... (Witten and Frank, 2005:126) Machine Learning for Language Technology 16(55)
- 17. Linear Classifiers Perceptron Choose a w that minimizes error w = arg min w t 1 − 1[yt = arg max y w · f(xt, y)] 1[p] = 1 p is true 0 otherwise This is a 0-1 loss function Aside: when minimizing error people tend to use hinge-loss or other smoother loss functions Machine Learning for Language Technology 17(55)
- 18. Linear Classifiers Perceptron Learning Algorithm Training data: T = {(xt, yt)} |T | t=1 1. w(0) = 0; i = 0 2. for n : 1..N 3. for t : 1..T (random order: shuffle the training set beforehand!) 4. Let y = arg maxy w(i) · f(xt, y) 5. if y = yt 6. w(i+1) = w(i) + f(xt, yt) − f(xt, y ) 7. i = i + 1 8. return wi See also Daume’ (2012:38-41). Machine Learning for Language Technology 18(55)
- 19. Linear Classifiers Perceptron: Separability and Margin Given an training instance (xt, yt), define: ¯Yt = Y − {yt} i.e., ¯Yt is the set of incorrect labels for xt A training set T is separable with margin γ > 0 if there exists a vector w with w = 1 such that: w · f(xt, yt) − w · f(xt, y ) ≥ γ for all y ∈ ¯Yt and ||w|| = j w2 j (Euclidean or L2 norm) Assumption: the training set is separable with margin γ Machine Learning for Language Technology 19(55)
- 20. Linear Classifiers Perceptron: Main Theorem Theorem: For any training set separable with a margin of γ, the following holds for the perceptron algorithm: mistakes made during training ≤ R2 γ2 where R ≥ ||f(xt, yt) − f(xt, y )|| for all (xt, yt) ∈ T and y ∈ ¯Yt Thus, after a finite number of training iterations, the error on the training set will converge to zero For proof, see the Appendix to these slides Machine Learning for Language Technology 20(55)
- 21. Linear Classifiers Perceptron Summary Learns a linear classifier that minimizes error Guaranteed to find a w in a finite amount of time Perceptron is an example of an online learning algorithm w is updated based on a single training instance in isolation w(i+1) = w(i) + f(xt, yt) − f(xt, y ) Compare decision trees that perform batch learning All training instances are used to find best split Machine Learning for Language Technology 21(55)
- 22. Linear Classifiers Margin Machine Learning for Language Technology 22(55)
- 23. Linear Classifiers Margin Training Testing Denote the value of the margin by γ Machine Learning for Language Technology 23(55)
- 24. Linear Classifiers Maximizing Margin (i) For a training set T , the margin of a weight vector w is the smallest γ such that w · f(xt, yt) − w · f(xt, y ) ≥ γ for every training instance (xt, yt) ∈ T , y ∈ ¯Yt Machine Learning for Language Technology 24(55)
- 25. Linear Classifiers Maximizing Margin (ii) Intuitively maximizing margin makes sense More importantly, generalization error to unseen test data is proportional to the inverse of the margin (for the proof, see Daume’, 2012: 45-46) ∝ R2 γ2 × |T | Perceptron: we have shown that: If a training set is separable by some margin, the perceptron will find a w that separates the data However, the perceptron does not pick w to maximize the margin! Machine Learning for Language Technology 25(55)
- 26. Linear Classifiers Maximizing Margin (iii) Let γ > 0 max ||w||≤1 γ such that: w · f(xt, yt) − w · f(xt, y ) ≥ γ ∀(xt, yt) ∈ T and y ∈ ¯Yt Note: algorithm still minimizes error ||w|| is bound since scaling trivially produces larger margin β(w · f(xt, yt) − w · f(xt, y )) ≥ βγ, for some β ≥ 1 Machine Learning for Language Technology 26(55)
- 27. Linear Classifiers Max Margin = Min Norm Let γ > 0 Max Margin: max ||w||≤1 γ such that: w·f(xt, yt)−w·f(xt, y ) ≥ γ ∀(xt, yt) ∈ T and y ∈ ¯Yt = Min Norm: min w 1 2 ||w||2 such that: w·f(xt, yt)−w·f(xt, y ) ≥ 1 ∀(xt, yt) ∈ T and y ∈ ¯Yt Instead of fixing ||w|| we fix the margin γ = 1 Technically γ ∝ 1/||w|| Machine Learning for Language Technology 27(55)
- 28. Linear Classifiers Support Vector Machines a.k.a. SVM(s) Machine Learning for Language Technology 28(55)
- 29. Linear Classifiers Support Vector Machines (i) min 1 2 ||w||2 such that: w · f(xt, yt) − w · f(xt, y ) ≥ 1 ∀(xt, yt) ∈ T and y ∈ ¯Yt Quadratic programming problem – a well known convex optimization problem Can be solved with out-of-the-box algorithms Batch learning algorithm – w set w.r.t. all training points Machine Learning for Language Technology 29(55)
- 30. Linear Classifiers Support Vector Machines (ii) Problem: Sometimes |T | is far too large Thus the number of constraints might make solving the quadratic programming problem very difficult Common technique: Sequential Minimal Optimization (SMO) Sparse: solution depends only on features in support vectors Machine Learning for Language Technology 30(55)
- 31. Linear Classifiers MIRA Margin Infused Relaxed Algorithm Machine Learning for Language Technology 31(55)
- 32. Linear Classifiers Margin Infused Relaxed Algorithm (MIRA) Another option – maximize margin using an online algorithm Batch vs. Online Batch – update parameters based on entire training set (SVM) Online – update parameters based on a single training instance at a time (Perceptron) MIRA can be thought of as a max-margin perceptron or an online SVM Machine Learning for Language Technology 32(55)
- 33. Linear Classifiers MIRA Batch (SVMs): min 1 2 ||w||2 such that: w·f(xt, yt)−w·f(xt, y ) ≥ 1 ∀(xt, yt) ∈ T and y ∈ ¯Yt Online (MIRA): Training data: T = {(xt, yt)} |T | t=1 1. w(0) = 0; i = 0 2. for n : 1..N 3. for t : 1..T 4. w(i+1) = arg minw* w* − w(i) such that: w · f(xt, yt) − w · f(xt, y ) ≥ 1 ∀y ∈ ¯Yt 5. i = i + 1 6. return wi MIRA has much smaller optimizations with only | ¯Yt| constraints Cost: sub-optimal optimization Machine Learning for Language Technology 33(55)
- 34. Linear Classifiers Interim Summary What we have covered Linear classifiers: Perceptron SVMs MIRA All are trained to minimize error With or without maximizing margin Online or batch What is next Logistic Regression Train linear classifiers to maximize likelihood Machine Learning for Language Technology 34(55)
- 35. Linear Classifiers Logistic Regression Machine Learning for Language Technology 35(55)
- 36. Linear Classifiers Logistic Regression (i) Define a conditional probability: P(y|x) = ew·f(x,y) Zx , where Zx = y ∈Y ew·f(x,y ) Note: still a linear classifier arg max y P(y|x) = arg max y ew·f(x,y) Zx = arg max y ew·f(x,y) = arg max y w · f(x, y) Machine Learning for Language Technology 36(55)
- 37. Linear Classifiers Logistic Regression (ii) P(y|x) = ew·f(x,y) Zx Q: How do we learn weights w A: Set weights to maximize log-likelihood of training data: w = arg max w t P(yt|xt) = arg max w t log P(yt|xt) In a nut shell we set the weights w so that we assign as much probability to the correct label y for each x in the training set Machine Learning for Language Technology 37(55)
- 38. Linear Classifiers Logistic Regression P(y|x) = ew·f(x,y) Zx , where Zx = y ∈Y ew·f(x,y ) w = arg max w t log P(yt|xt) (*) The objective function (*) is concave Therefore there is a global maximum No closed form solution, but lots of numerical techniques Gradient methods (gradient ascent, iterative scaling) Newton methods (limited-memory quasi-newton) Machine Learning for Language Technology 38(55)
- 39. Linear Classifiers Logistic Regression Summary Define conditional probability P(y|x) = ew·f(x,y) Zx Set weights to maximize log-likelihood of training data: w = arg max w t log P(yt|xt) Can find the gradient and run gradient ascent (or any gradient-based optimization algorithm) F(w) = ( ∂ ∂w0 F(w), ∂ ∂w1 F(w), . . . , ∂ ∂wm F(w)) ∂ ∂wi F(w) = t fi (xt, yt) − t y ∈Y P(y |xt)fi (xt, y ) Machine Learning for Language Technology 39(55)
- 40. Linear Classifiers Linear Classification: Summary Basic form of (multiclass) classifier: y = arg max y w · f(x, y) Different learning methods: Perceptron – separate data (0-1 loss, online) Support vector machine – maximize margin (hinge loss, batch) Logistic regression – maximize likelihood (log loss, batch) All three methods are widely used in NLP Machine Learning for Language Technology 40(55)
- 41. Linear Classifiers Aside: Min error versus max log-likelihood (i) Highly related but not identical Example: consider a training set T with 1001 points 1000 × (xi , y = 0) = [−1, 1, 0, 0] for i = 1 . . . 1000 1 × (x1001, y = 1) = [0, 0, 3, 1] Now consider w = [−1, 0, 1, 0] Error in this case is 0 – so w minimizes error [−1, 0, 1, 0] · [−1, 1, 0, 0] = 1 > [−1, 0, 1, 0] · [0, 0, −1, 1] = −1 [−1, 0, 1, 0] · [0, 0, 3, 1] = 3 > [−1, 0, 1, 0] · [3, 1, 0, 0] = −3 However, log-likelihood = −126.9 (omit calculation) Machine Learning for Language Technology 41(55)
- 42. Linear Classifiers Aside: Min error versus max log-likelihood (ii) Highly related but not identical Example: consider a training set T with 1001 points 1000 × (xi , y = 0) = [−1, 1, 0, 0] for i = 1 . . . 1000 1 × (x1001, y = 1) = [0, 0, 3, 1] Now consider w = [−1, 7, 1, 0] Error in this case is 1 – so w does not minimizes error [−1, 7, 1, 0] · [−1, 1, 0, 0] = 8 > [−1, 7, 1, 0] · [0, 0, −1, 1] = −1 [−1, 7, 1, 0] · [0, 0, 3, 1] = 3 < [−1, 7, 1, 0] · [3, 1, 0, 0] = 4 However, log-likelihood = -1.4 Better log-likelihood and worse error Machine Learning for Language Technology 42(55)
- 43. Linear Classifiers Aside: Min error versus max log-likelihood (iii) Max likelihood = min error Max likelihood pushes as much probability on correct labeling of training instance Even at the cost of mislabeling a few examples Min error forces all training instances to be correctly classified SVMs with slack variables – allows some examples to be classified wrong if resulting margin is improved on other examples Machine Learning for Language Technology 43(55)
- 44. Linear Classifiers Aside: Max margin versus max log-likelihood Let’s re-write the max likelihood objective function w = arg max w t log P(yt|xt) = arg max w t log ew·f(xt ,yt ) y ∈Y ew·f(x,y ) = arg max w t w · f(xt, yt) − log y ∈Y ew·f(x,y ) Pick w to maximize score difference between correct labeling and every possible labeling Margin: maximize difference between correct and all incorrect The above formulation is often referred to as the soft-margin Machine Learning for Language Technology 44(55)
- 45. Linear Classifiers Aside: Logistic Regression = Maximum Entropy Well known equivalence Max Ent: maximize entropy subject to constraints on features Empirical feature counts must equal expected counts Quick intuition Partial derivative in logistic regression ∂ ∂wi F(w) = t fi (xt, yt) − t y ∈Y P(y |xt)fi (xt, y ) First term is empirical feature counts and second term is expected counts Derivative set to zero maximizes function Therefore when both counts are equivalent, we optimize the logistic regression objective! Machine Learning for Language Technology 45(55)
- 46. Appendix Proofs and Derivations Machine Learning for Language Technology 46(55)
- 47. Convergence Proof for Perceptron Perceptron Learning Algorithm Training data: T = {(xt , yt )} |T | t=1 1. w(0) = 0; i = 0 2. for n : 1..N 3. for t : 1..T 4. Let y = arg maxy w(i) · f(xt , y) 5. if y = yt 6. w(i+1) = w(i) + f(xt , yt ) − f(xt , y ) 7. i = i + 1 8. return wi w(k−1) are the weights before kth mistake Suppose kth mistake made at the tth example, (xt , yt ) y = arg maxy w(k−1) · f(xt , y) y = yt w(k) = w(k−1) + f(xt , yt ) − f(xt , y ) Now: u · w(k) = u · w(k−1) + u · (f(xt , yt ) − f(xt , y )) ≥ u · w(k−1) + γ Now: w(0) = 0 and u · w(0) = 0, by induction on k, u · w(k) ≥ kγ Now: since u · w(k) ≤ ||u|| × ||w(k)|| and ||u|| = 1 then ||w(k)|| ≥ kγ Now: ||w(k) ||2 = ||w(k−1) ||2 + ||f(xt , yt ) − f(xt , y )||2 + 2w(k−1) · (f(xt , yt ) − f(xt , y )) ||w(k) ||2 ≤ ||w(k−1) ||2 + R2 (since R ≥ ||f(xt , yt ) − f(xt , y )|| and w(k−1) · f(xt , yt ) − w(k−1) · f(xt , y ) ≤ 0) Machine Learning for Language Technology 47(55)
- 48. Convergence Proof for Perceptron Perceptron Learning Algorithm We have just shown that ||w(k)|| ≥ kγ and ||w(k)||2 ≤ ||w(k−1)||2 + R2 By induction on k and since w(0) = 0 and ||w(0)||2 = 0 ||w(k) ||2 ≤ kR2 Therefore, k2 γ2 ≤ ||w(k) ||2 ≤ kR2 and solving for k k ≤ R2 γ2 Therefore the number of errors is bounded! Machine Learning for Language Technology 48(55)
- 49. Gradient Ascent for Logistic Regression Gradient Ascent Let F(w) = t log ew·f(xt ,yt ) Zx Want to find arg maxw F(w) Set w0 = Om Iterate until convergence wi = wi−1 + α F(wi−1 ) α > 0 and set so that F(wi ) > F(wi−1) F(w) is gradient of F w.r.t. w A gradient is all partial derivatives over variables wi i.e., F(w) = ( ∂ ∂w0 F(w), ∂ ∂w1 F(w), . . . , ∂ ∂wm F(w)) Gradient ascent will always find w to maximize F Machine Learning for Language Technology 49(55)
- 50. Gradient Ascent for Logistic Regression The partial derivatives Need to find all partial derivatives ∂ ∂wi F(w) F(w) = t log P(yt|xt) = t log ew·f(xt ,yt ) y ∈Y ew·f(xt ,y ) = t log e P j wj ×fj (xt ,yt ) y ∈Y e P j wj ×fj (xt ,y ) Machine Learning for Language Technology 50(55)
- 51. Gradient Ascent for Logistic Regression Partial derivatives - some reminders 1. ∂ ∂x log F = 1 F ∂ ∂x F We always assume log is the natural logarithm loge 2. ∂ ∂x eF = eF ∂ ∂x F 3. ∂ ∂x t Ft = t ∂ ∂x Ft 4. ∂ ∂x F G = G ∂ ∂x F−F ∂ ∂x G G2 Machine Learning for Language Technology 51(55)
- 52. Gradient Ascent for Logistic Regression The partial derivatives ∂ ∂wi F(w) = ∂ ∂wi t log e P j wj ×fj (xt ,yt ) y ∈Y e P j wj ×fj (xt ,y ) = t ∂ ∂wi log e P j wj ×fj (xt ,yt ) y ∈Y e P j wj ×fj (xt ,y ) = t ( y ∈Y e P j wj ×fj (xt ,y ) e P j wj ×fj (xt ,yt ) )( ∂ ∂wi e P j wj ×fj (xt ,yt ) y ∈Y e P wj wj ×fj (xt ,y ) ) = t ( Zxt e P j wj ×fj (xt ,yt ) )( ∂ ∂wi e P j wj ×fj (xt ,yt ) Zxt ) Machine Learning for Language Technology 52(55)
- 53. Gradient Ascent for Logistic Regression The partial derivatives Now, ∂ ∂wi e P j wj ×fj (xt ,yt ) Zxt = Zxt ∂ ∂wi e P j wj ×fj (xt ,yt ) − e P j wj ×fj (xt ,yt ) ∂ ∂wi Zxt Z2 xt = Zxt e P j wj ×fj (xt ,yt ) fi (xt , yt ) − e P j wj ×fj (xt ,yt ) ∂ ∂wi Zxt Z2 xt = e P j wj ×fj (xt ,yt ) Z2 xt (Zxt fi (xt , yt ) − ∂ ∂wi Zxt ) = e P j wj ×fj (xt ,yt ) Z2 xt (Zxt fi (xt , yt ) − X y ∈Y e P j wj ×fj (xt ,y ) fi (xt , y )) because ∂ ∂wi Zxt = ∂ ∂wi X y ∈Y e P j wj ×fj (xt ,y ) = X y ∈Y e P j wj ×fj (xt ,y ) fi (xt , y ) Machine Learning for Language Technology 53(55)
- 54. Gradient Ascent for Logistic Regression The partial derivatives From before, ∂ ∂wi e P j wj ×fj (xt ,yt ) Zxt = e P j wj ×fj (xt ,yt ) Z2 xt (Zxt fi (xt , yt ) − X y ∈Y e P j wj ×fj (xt ,y ) fi (xt , y )) Sub this in, ∂ ∂wi F(w) = X t ( Zxt e P j wj ×fj (xt ,yt ) )( ∂ ∂wi e P j wj ×fj (xt ,yt ) Zxt ) = X t 1 Zxt (Zxt fi (xt , yt ) − X y ∈Y e P j wj ×fj (xt ,y ) fi (xt , y ))) = X t fi (xt , yt ) − X t X y ∈Y e P j wj ×fj (xt ,y ) Zxt fi (xt , y ) = X t fi (xt , yt ) − X t X y ∈Y P(y |xt )fi (xt , y ) Machine Learning for Language Technology 54(55)
- 55. Gradient Ascent for Logistic Regression FINALLY!!! After all that, ∂ ∂wi F(w) = t fi (xt, yt) − t y ∈Y P(y |xt)fi (xt, y ) And the gradient is: F(w) = ( ∂ ∂w0 F(w), ∂ ∂w1 F(w), . . . , ∂ ∂wm F(w)) So we can now use gradient assent to find w!! Machine Learning for Language Technology 55(55)