Perceptrons and Linear Classifiers William Cohen 2-4-2008

Perceptrons and Linear Classifiers

William Cohen

2-4-2008

Announcement: no office hours for William this Friday 2/8

Dave Touretzky’s Gallery of CSS Descramblers

Linear Classifiers

• Let’s simplify life by assuming:– Every instance is a vector of real numbers, x=(x1,…,xn).

(Notation: boldface x is a vector.)– There are only two classes, y=(+1) and y=(-1)

• A linear classifier is vector w of the same dimension as x that is used to make this prediction:

)(sign)...sign(ˆ 2211 xw nnxwxwxwy

0if1

0 if1)(sign

xx

w

-W

Visually, x · w is the distance you get if you “project x onto w”

X1

x2X1

. w

X2 . w

The line perpendicular to w divides the vectors classified as positive from the vectors classified as negative.

In 3d: lineplaneIn 4d: planehyperplane…

w

-W

)(sign)...sign(ˆ 2211 xw nnxwxwxwy

Wolfram MathWorld

Mediaboost.comGeocities.com/bharatvarsha1947

w

-W

Notice that the separating hyperplane goes through the origin…if we don’t want this we can preprocess our examples:

nxxx ,...,, 21x

nxxx ,...,,,1 21x

)(sign)...sign(ˆ 2211 xw nnxwxwxwy

)(sign)...1sign(ˆ 22110 xw nnxwxwxwwy

What have we given up?,...,,,,,,

,....,

,,,,,,

1

cooltempmildtemphottemprainoutlookovercastoutlooksunnyoutlook

n

xxxxxx

xx

-1

+1

1,0 0,1, ,1,0,0 ,0,1,0

,...,0,1,07 ,,,

rainoutlookovercastoutlooksunnyoutlook xxxD

Outlook overcast Humidity normal

What have we given up?

• Not much!– Practically, it’s a little harder to understand a

particular example (or classifier)– Practically, it’s a little harder to debug

• You can still express the same information• You can analyze things mathematically much

more easily

Naïve Bayes as a Linear Classifier

Consider Naïve Bayes with two classes (+1, -1) and binary features (0,1).

Naïve Bayes as a Linear Classifier

Naïve Bayes as a Linear Classifier

“log odds”

Naïve Bayes as a Linear Classifierpi

qi

Naïve Bayes as a Linear Classifier

Naïve Bayes as a Linear Classifier

• Summary: – NB is linear classifier

•

– Weights wi have a closed form• which is fairly simple, expressed in log-odds

)(signˆ xwy

Proceedings of ECML-98,

10th European Conference on Machine Learning

An Even Older Linear Classifier

• 1957: The perceptron algorithm: Rosenblatt – WP: “A handsome bachelor, he drove a classic MGA sports car and was often seen with his

cat named Tobermory. He enjoyed mixing with undergraduates, and for several years taught an interdisciplinary undergraduate honors course entitled "Theory of Brain Mechanisms" that drew students equally from Cornell's Engineering and Liberal Arts colleges…this course was a melange of ideas .. experimental brain surgery on epileptic patients while conscious, experiments on .. the visual cortex of cats, ... analog and digital electronic circuits that modeled various details of neuronal behavior (i.e. the perceptron itself, as a machine).”

– Built on work of Hebbs (1949); also developed by Widrow-Hoff (1960)

• 1960: Perceptron Mark 1 Computer – hardware implementation

Bell Labs TM 59-1142-11– Datamation 1961 – April 1 1984 Special Edition of CACM

An Even Older Linear Classifier

• 1957: The perceptron algorithm: Rosenblatt – WP: “A handsome bachelor, he drove a classic MGA sports car and was often seen with his

cat named Tobermory. He enjoyed mixing with undergraduates, and for several years taught an interdisciplinary undergraduate honors course entitled "Theory of Brain Mechanisms" that drew students equally from Cornell's Engineering and Liberal Arts colleges…this course was a melange of ideas .. experimental brain surgery on epileptic patients while conscious, experiments on .. the visual cortex of cats, ... analog and digital electronic circuits that modeled various details of neuronal behavior (i.e. the perceptron itself, as a machine).”

– Built on work of Hebbs (1949); also developed by Widrow-Hoff (1960)

• 1960: Perceptron Mark 1 Computer – hardware implementation• 1969: Minksky & Papert book shows perceptrons limited to linearly

separable data, and Rosenblatt dies in boating accident• 1970’s: learning methods for two-layer neural networks• Mid-late 1980’s (Littlestone & Warmuth): mistake-bounded learning

& analysis of Winnow method; early-mid 1990’s, analyses of perceptron/Widrow-Hoff

Experimental evaluation of Perceptron vs WH and Experts (Winnow-like methods) in SIGIR-1996 (Lewis,

Schapire, Callan, Papka), and (Cohen & Singer)

Freund & Schapire, 1998-1999 showed “kernel trick” and averaging/voting worked

The voted perceptron

A Binstance xi Compute: yi = sign(vk . xi )

^

yi

^

yi

If mistake: vk+1 = vk + yi xi

u

-u

2γ

u

-u

2γ

+x1v1

(1) A target u (2) The guess v1 after one positive example.

u

-u

2γ

u

-u

2γ

v1

+x2

v2

+x1v1

-x2

v2

(3a) The guess v2 after the two positive examples: v2=v1+x2

(3b) The guess v2 after the one positive and one negative example: v2=v1-x2

I want to show two things:

1. The v’s get closer and closer to u: v.u increases with each mistake

2. The v’s do not get too large: v.v grows slowly

u

-u

2γ

u

-u

2γ

v1

+x2

v2

+x1v1

-x2

v2

(3a) The guess v2 after the two positive examples: v2=v1+x2

(3b) The guess v2 after the one positive and one negative example: v2=v1-x2

> γ

u

-u

2γ

u

-u

2γ

v1

+x2

v2

+x1v1

-x2

v2

On-line to batch learning

1. Pick a vk at random according to mk/m, the fraction of examples it was used for.

2. Predict using the vk you just picked.

3. (Actually, use some sort of deterministic approximation to this).

The voted perceptron

Some more commentsPerceptrons are like support vector machines (SVMs)

1. SVMs search for something that looks like u: i.e., a vector w where ||w|| is small and the margin for every example is large

2. You can use “the kernel trick” with perceptrons• Replace x.w with (x.w+1)d

Experimental Results

Task: classifying hand-written digits for the post office

More Experimental Results (Linear kernel, one pass over the data)