I was at this talk and it was appallingly bad. The most serious confusion is that the algorithms behind the prediction API are NOT the same as the algorithms described in the talk. The talk was really two talks glued together without a transition. The first part was essentially just a rehash of the Prediction API press release. The second part was a debased pastiche of two machine learning papers by Googlers that showed speedup on a few problems using fairly well-known approaches.
See http://arxiv.org/abs/0911.0491 for a counterpoint that describes why delay is bad in learning. See here http://leon.bottou.org/projects/sgd for some discussion on why SGD is hard to beat in parallel techniques. Regardless of all that, it would be interesting to provide an alternative service. On Fri, Aug 6, 2010 at 11:02 AM, Grant Ingersoll <[email protected]>wrote: > Just had this forwarded to me: > http://docs.google.com/a/m4xl1n.com/present/view?id=0AQ85RzjPG_zXZHFoZzhkNl8ybm1wamhkZmc&hl=en > > Especially interesting, since we implement a number of the classifiers > mentioned. Now, just need a web service and seamless EC2 integration... > > -Grant
