Distributed Discriminative machine learning

Discriminative parameter estimation involves updating a weight vector. For really big problems, this cannot be done on one machine. However, it is possible to decompose the problem and solve it using Map Reduce.

This project will use large-scale labelled data sets. An example problem is case restoration: predict whether some word should be upper or lower case. (This is important for machine translation). The labelled training set would be a large corpus of data, with each word labelled as being either uppoer or lower case.

A starting point on using Map Reduce for training a perceptron