|
Hello, I'm currently working on a two step model that both does a feature reduction (Factor Analysis, PCA, K-means, etc) and then does a binary classification over the newly reduced features and tests it in a different domain (Domain adaptation). I was wondering if there are any models that do both steps in one swoop and as a single inference procedure, if there's not, it might be a good idea for a paper. Thanks |
|
I think this paper does basically the training procedure you're talking about with k-means, but I don't think they evaluate it for domain adaptation: http://www.kyb.mpg.de/publications/attachments/NIPS2007-Ye_4710[0].pdf Here's one for PCA: http://www.wseas.us/e-library/conferences/2006bucharest/papers/518-433.pdf Here's a paper some of my friends wrote using autoencoders for domain adaptation. Since autoencoders are gradient-based you could try re-implementing their model but training the whole thing jointly instead of training it one piece at a time: eprints.pascal-network.org/archive/00008597/01/342_icmlpaper.pdf |
That sounds like a 3 layer neural net with hidden layer less than input layer
marshallp, how would you fit the domain adaptation bit in?
I don't really get the point of domain adaptation. Everything should just be part of the training set. So the model doesn't perform well in the new domain, just train a new model with the extra data, or run further training on the existing model with the new data, or create an ensemble by adding a new model etc. Just do all of those in parallel automatically and choose the best performer.