Handwritten digit classification using logistic regression and support vector machines
The repo contains implementation of various models to classify handwritten digits.
-
Model 1 : Binary logistic regression
-> In this model ten one-vs-all classifier corresponding to each digit are built.
-> During prediction the output of classifier with maximum value is considered.
-
Model 2 : Multiclass logistic regression
-> This model is an extension of one-vs-all classifier.
-> In this case the classifier outputs the K probabilities each representing the chance of input belonging to one of the K classes.
I followed the text in Bishop's book (pg 205-210) to implement the multi class logistic regression.
NOTE : In Newton Raphson's method where we compute the Hessian might throw out of memory error. Make sure you change the default memory setting of you Matlab installation/setup before running.
-
Support Vector machines
-> libsvm is being used, no explicit code.