Hacker News new | past | comments | ask | show | jobs | submit login

This is related to last year's article about Google's neural network being able to recognize cats, right?

http://www.nytimes.com/2012/06/26/technology/in-a-big-networ...




I don't think so. The neural network in Google+ was trained on labeled images and now finds similar objects in unlabeled images.

The technology discussed in that article is about deducing the existence of a common feature, in this instance a cat, from a large collection of unlabelled images.


It may be the same tech (roughly). Use the same approach for all but the last layer, then use traditional backprop to learn the last layer and fine tune the connections in the lower layers.

Mostly unlabelled then, which means you can learn to generalise over a huge number of images but learn labels on a smaller set.


Yep. I don't know if that's what is actually being used here, but that is pretty much how they did it with the same system:

"We applied the feature learning method to the task of recognizing objects in the ImageNet dataset (Deng et al., 2009). After unsupervised training on YouTube and ImageNet images, we added one-versus-all logistic classifiers on top of the highest layer. We first trained the logistic classifiers and then fine-tuned the network. Regularization was not employed in the logistic classifiers. The entire training was carried out on 2,000 machines for one week."[1]

Basically you learn features in unlabeled data, then identify the features your trained net is recognizing with labeled data. When you run over g+ images, you then only tag with features you're sure of past some threshold of certainty.

[1] http://research.google.com/pubs/pub38115.html


I thought the training images were unlabeled. See the actual paper: http://research.google.com/pubs/pub38115.html

"Abstract: We consider the problem of building highlevel, class-specific feature detectors from only unlabeled data."




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: