Feb 2, 2012 · 7:00 PM
This location is shown only to members
We're really excited for our next event with Joseph Turian! Come at 6:30 to network with the group.
ABSTRACT:This talk will present a handful of recipes for cheaply improving the accuracy of existing models.
Sure, if you want more accuracy, you can read the latest papers and spend a bunch of time implementing new fangled models. It's definitely fun to brag about how you took the long road. But what if you don't have the time to invest in something nebulous and tricky, and simply want to solve your problem?
A simpler approach is to take your existing model, and shove more data in. You can annotate more data, with active learning. You can insert features from an unsupervised model. You can use distant supervision. By presenting a handful of recipes and examples, this talk will encourage you to think outside the box, and figure out how to use more data to improve existing models.
This talk is the opposite of my O'Reilly Strata Conference talk "New Developments in Large Data Techniques". That talk was about awesome new models that no one in industry is using yet:http://www.youtube.com/watch?v=fEUw8igr1IY
Joseph Turian, Ph.D., is the founder of MetaOptimize, a consultancy onML and NLP.