Machine Learning Basics Lecture 3: Perceptron. Princeton University COS 495 Instructor: Yingyu Liang

Transcription

1 Machine Learning Basics Lecture 3: Perceptron Princeton University COS 495 Instructor: Yingyu Liang

2 Perceptron

3 Overview Previous lectures: (Principle for loss function) MLE to derive loss Example: linear regression; some linear classification models This lecture: (Principle for optimization) local improvement Example: Perceptron; SGD

4 Task (w ) T x = 0 (w ) T x > 0 Class +1 w (w ) T x < 0 Class -1

5 Attempt Given training data x i, y i : 1 i n i.i.d. from distribution D Hypothesis f w x y = +1 if w T x > 0 y = 1 if w T x < 0 = w T x Prediction: y = sign(f w x ) = sign(w T x) Goal: minimize classification error

6 Perceptron Algorithm Assume for simplicity: all x i has length 1 Perceptron: figure from the lecture note of Nina Balcan

7 Intuition: correct the current mistake If mistake on a positive example w T t+1 x = w t + x T x = w T t x + x T x = w T t x + 1 If mistake on a negative example w T t+1 x = w t x T x = w T t x x T x = w T t x 1

8 The Perceptron Theorem Suppose there exists w that correctly classifies x i, y i W.L.O.G., all x i and w have length 1, so the minimum distance of any example to the decision boundary is γ = min i w T x i Then Perceptron makes at most 1 γ 2 mistakes

9 The Perceptron Theorem Suppose there exists w that correctly classifies x i, y i W.L.O.G., all x i and w have length 1, so the minimum distance of any example to the decision boundary is γ = min i w T x i Need not be i.i.d.! Then Perceptron makes at most 1 γ 2 mistakes Do not depend on n, the length of the data sequence!

10 Analysis First look at the quantity w t T w Claim 1: w T t+1 w w T t w + γ Proof: If mistake on a positive example x w T t+1 w = w t + x T w = w T t w + x T w w T t w + γ If mistake on a negative example w T t+1 w = w t x T w = w T t w x T w w T t w + γ

11 Analysis Next look at the quantity w t Negative since we made a mistake on x Claim 2: w t+1 2 wt Proof: If mistake on a positive example x w t+1 2 = wt + x 2 = w t 2 + x 2 + 2wt T x

12 Analysis: putting things together Claim 1: w T t+1 w w T t w + γ 2 2 Claim 2: w t+1 wt + 1 After M mistakes: T w M+1 w γm w M+1 M T w M+1 w w M+1 So γm M, and thus M 1 γ 2

13 Intuition Claim 1: w T t+1 w w T t w + γ Claim 2: w t+1 2 wt The correlation gets larger. Could be: 1. w t+1 gets closer to w 2. w t+1 gets much longer Rules out the bad case 2. w t+1 gets much longer

14 Some side notes on Perceptron

15 History Figure from Pattern Recognition and Machine Learning, Bishop

16 Note: connectionism vs symbolism Symbolism: AI can be achieved by representing concepts as symbols Example: rule-based expert system, formal grammar Connectionism: explain intellectual abilities using connections between neurons (i.e., artificial neural networks) Example: perceptron, larger scale neural networks

17 Symbolism example: Credit Risk Analysis Example from Machine learning lecture notes by Tom Mitchell

18 Connectionism example Neuron/perceptron Figure from Pattern Recognition and machine learning, Bishop

19 Note: connectionism v.s. symbolism Formal theories of logical reasoning, grammar, and other higher mental faculties compel us to think of the mind as a machine for rulebased manipulation of highly structured arrays of symbols. What we know of the brain compels us to think of human information processing in terms of manipulation of a large unstructured set of numbers, the activity levels of interconnected neurons The Central Paradox of Cognition (Smolensky et al., 1992)

20 Note: online vs batch Batch: Given training data x i, y i : 1 i n, typically i.i.d. Online: data points arrive one by one 1. The algorithm receives an unlabeled example x i 2. The algorithm predicts a classification of this example. 3. The algorithm is then told the correct answer y i, and update its model

21 Stochastic gradient descent (SGD)

22 Gradient descent Minimize loss L θ, where the hypothesis is parametrized by θ Gradient descent Initialize θ 0 θ t+1 = θ t η t L θ t

23 Stochastic gradient descent (SGD) Suppose data points arrive one by one L θ = 1 σ n n t=1 l(θ, x t, y t ), but we only know l(θ, x t, y t ) at time t Idea: simply do what you can based on local information Initialize θ 0 θ t+1 = θ t η t l(θ t, x t, y t )

24 Example 1: linear regression Find f w x = w T x that minimizes L f w = 1 σ n n t=1 w T x t y 2 t l w, x t, y t = 1 n wt x t y t 2 w t+1 = w t η t l w t, x t, y t = w t 2η t n w t T x t y t x t

25 Example 2: logistic regression Find w that minimizes L w = 1 n yt=1 logσ(w T x t ) 1 n log[1 σ w T x t ] y t = 1 L w = 1 n t logσ(y t w T x t ) l w, x t, y t = 1 n logσ(y tw T x t )

26 Example 2: logistic regression Find w that minimizes l w, x t, y t = 1 n logσ(y tw T x t ) w t+1 = w t η t l w t, x t, y t = w t + η t n σ a 1 σ a σ(a) y t x t Where a = y t w t T x t

27 Example 3: Perceptron Hypothesis: y = sign(w T x) Define hinge loss l w, x t, y t = y t w T x t I[mistake on x t ] L w = t y t w T x t I[mistake on x t ] w t+1 = w t η t l w t, x t, y t = w t + η t y t x t I[mistake on x t ]

28 Example 3: Perceptron Hypothesis: y = sign(w T x) w t+1 = w t η t l w t, x t, y t = w t + η t y t x t I[mistake on x t ] Set η t = 1. If mistake on a positive example w t+1 = w t + y t x t = w t + x If mistake on a negative example w t+1 = w t + y t x t = w t x

29 Pros & Cons Pros: Widely applicable Easy to implement in most cases Guarantees for many losses Good performance: error/running time/memory etc. Cons: No guarantees for non-convex opt (e.g., those in deep learning) Hyper-parameters: initialization, learning rate

30 Mini-batch Instead of one data point, work with a small batch of b points (x tb+1, y tb+1 ),, (x tb+b, y tb+b ) Update rule θ t+1 = θ t η t 1 b 1 i b l θ t, x tb+i, y tb+i Other variants: variance reduction etc.

31 Homework

32 Homework 1 Assignment online Course website: Piazza: Due date: Feb 17 th (one week) Submission Math part: hand-written/print; submit to TA (Office: EE, C319B) Coding part: in Matlab/Python; submit the.m/.py file on Piazza

33 Homework 1 Grading policy: every late day reduces the attainable credit for the exercise by 10%. Collaboration: Discussion on the problem sets is allowed Students are expected to finish the homework by himself/herself The people you discussed with on assignments should be clearly detailed: before the solution to each question, list all people that you discussed with on that particular question.