Volcano Pucaya

John Langford

Doctor of Learning at Microsoft Research
{jl@hunch.net, jcl@microsoft.com}
pgp key
Vowpal Wabbit Fast Learning Project
Learning to Interact tutorial at NIPS
Extreme Multiclass workshop @ NIPS
Machine Learning (Theory) Weblog Research Conferences, Workshops, Seminars, Summer Schools Classes Resume

Start Page Yan Essays Interesting Software Linux Designs The Doghouse

I am a technology maker, focusing on learning.

Interactive Learning

This is all about learning when the learning algorithm has some interaction with the environment.


Reductions from one learning problem to another create meta-algorithms reusing old algorithms. This method can be analyzed without assumptions and often yields superior performance.

Hashing for Learning

This is all about how to compress very large learning algorithms into much smaller space than you might have believed possible using a hashing trick.

Cover Tree

A cover tree is a datastructure for fast general proximity queries with both theoretical guarantees and state-of-the-art practical performance.

Tight Bounds

How can you measure prediction ability? I created refined general bounds with applications for learning measurement and benchmark design.


Captchas block machines (but not humans) from data and create concrete challenges for AI and learning.


Isomap is an algorithm for extracting low dimensional manifolds from points in a high dimensional space.
More research.

My old webpage (with some other content).
A labelled image dataset from the espgame. This dataset can be searched.
The peekaboom dataset.