how to get base vector from deep learning

User 1209 | 1/17/2015, 4:24:08 AM

can any experts here tell me: is it possible to get the learned base vectors from each layer and visualize them? if yes, how? thanks


User 940 | 1/19/2015, 9:44:06 PM

Hi Charley,

Could you clarify what you mean by 'learned base vector'?

Currently we support using a deep neural network as a feature extractor. This works by propagating an input through the network until a specified layer, at which point the activations are placed in a vector. This is the extracted feature vector for that input. This can be visualized with GraphLab Canvas, or with any other tool you would like.

For more information, look at the <a href="">API docs</a>

Let us know if that answered your question!

Cheers! -Piotr

User 1209 | 1/19/2015, 10:21:54 PM

I mean, for example, we decompose vec(x) = 0.8vec(b1)+ 0.2vec(b2)+... here I guess the coefficients 0.8, 0.2 are the feature vector you are talking about, but I also want to see b1, b2, which are learned edges from images. For example, if we talk about elephant, the lower level base vector would be some edges, higher level could be the legs....

Is this possible to get them in create? It's obviously there, but just not sure you have the API to retrieve it? Or I was wrong?

User 940 | 1/20/2015, 1:19:46 AM

Hi Charley,

Unfortunately, you are correct in that we do not support visualization of what the network is learning. However, we always take feature requests. Is this something that would be useful to you, and in what way?

Cheers! -Piotr

User 1209 | 1/20/2015, 3:19:29 PM

hi Piotr, thanks for your answers. This feature is useful because it allows we understand more about we are learning behind the scene, instead of treating it as a total black box. They will be interesting to non-technical persons especially to be more confident about the technology. If it is not too hard to retrieve them in the existing framework, it will be great to have them, my 2c.