perceptron dichotomizer

Download Report

Transcript perceptron dichotomizer

Multivariate linear models for
regression and classification
Outline:
1) multivariate linear regression
2) linear classification (perceptron)
3) logistic regression
Multivariate Linear Classification
The perceptron
(lecture 3 on amlbook.com)
Review: perceptron applied to credit approval
Integrate “threshold” into input vector
Perceptron learning algorithm (PLA)
sign(wTx) negative
sign(wTx) positive
Each iteration pulls boundary in direction that tends to correct misclassification
If data is linearly separable, iteration terminates when Ein(g) = 0
Otherwise terminate by maximum number of iterations
A graphical demo of the Perceptron Learning Algorithm
(in Octave) is on AMLbook.com
Classification for digit recognition
Examples of hand-written digits from zip codes
Given 16x16 pixel image, could have model with 257 weights
(attributes)
Better to develop a small number of attributes (features) from
the raw data
Regardless of attributes chosen, x0 = 1 is included
2-attribute model: intensity and symmetry
Intensity: how much black is in the image
Symmetry: how similar are mirror images
PLA is unstable. No obvious trend toward min Ein
Correcting misclassification of one example may cause others to
become misclassified
Pocket algorithm (text p80)
Keeps the best PLA result obtained thus far
Must evaluate Ein (number of misclassified examples) after
each update to determine if it is better than previous best
PLA & Pocket final results of 1 vs 5 classification
• Just like in Assignment 4, multivariate linear regression
can be used as an alternative to PLA and Pocket algorithm.
• PLA and Pocket algorithm require class labels to be +1.
• Multivariate linear regression can use any 2 distinct
real numbers r1 and r2.
Classification boundaries
• yfit (x) = w0 + w1 x1 + w2 x2
• yfit (x) = (r1 + r2)/2 defines the a function of x1 and x2
that is the boundary
• Solve this function for x2 as a function of x1
Assignment 5: due 10-21-14
Part 1:
Apply Pocket algorithm (text p80) to 1vs5 classification
Two attribute files (intensity, symmetry) are on class webpage.
Use the smaller (424 examples) for training. Use the larger
(1561 examples) get the following results:
1) Etest as percent misclassified
2) Confusion matrix as percentages
3) Scatter plot with boundary (as in slide 12)
Part2:
Compare Pocket algorithm results with those obtained by multivariate
linear regression. Again, use smaller set for training and larger for
testing. Report the same results as above. Calculate Etest as mean of
squared residuals by leave-one-out (424 runs). Compare to Etest as
mean of squared residuals calculated from the test set.