Semi-supervised learning

User 2334 | 4/12/2016, 7:02:42 AM

Hi!

Is it possible to use Dato for Semi-supervised learning? I have 30.000 labeled images and about 5 times more unlabeled images. I tried to use neural networks on the labeled images with extracting features from imagenet but I am not happy with my accuray.

Any suggestions? Thanks,

Comments

User 2593 | 4/12/2016, 9:54:38 PM

Hi @ete,

Yes you can definitely do something with GLC here. Here are a few suggestions:

  • Use our deep feature extractor to extract features from all your images. One the 30k labelled images, train a logistic classifier using the deep features as your independent variables. Use the model to then label your unlabelled data

  • Use our deep feature extractor to extract features from all your images. Build a similarity graph using our nearest neighbors toolkit. Now you can do label propagation on the graph.

Let me know if either of these work!

Charlie


User 2334 | 4/13/2016, 6:58:35 AM

Thanks for the suggestions Charlie! I will give it a shot and let you know whether my accuracy increased. Great tip!