Machine Learning

I was at the Intl Conf on Machine Learning (ICML) last week, in the workshop on Knowledge Discovery from Data Streams. I gave a tutorial on new developments in data stream algorithms. I am not a machine learning expert, but I spent some time formulating machine learning problems in the data stream model. For example, how to approximate a decision tree that is much too large for the storage available with some approximation guarantees, or how to find the support vector with some provable generalization error guarantee when the positive and negative examples arrive on a stream, etc. Still, it was a difficult conversation at the conference. A lot of researchers seem to think any incremental algorithm is a streaming algorithm. Sigh.