"Guided" machine learning?

(John Wu) #1

Does anybody know of a deep neural network approach that can be “guided” to predict quantity or category B after it has been trained to predict A? Let me describe an example:

I work in astrophysics and I’ve recently become interested in training convolutional networks on RGB images of galaxies in order to predict other quantities. I know that predicting morphological classification is quite possible based on a Kaggle competition and also from discussion with @farlion. However, I’d also like to predict with CNNs, for example, the chemical composition of galaxies, which can be determined from spectroscopic measurements taken with telescopes. In this case, the chemical abundance is quantity B and the morphology is quantity A.

Suppose that A correlates with B. I am curious if it would be worthwhile to “guide” the network to learn the abstraction of morphology (A) through the usual training procedure, and then to snip off the final activation layers and add a few more layers and finally predict the chemical abundance (B).

Does this idea sound plausible, and if so, has it been implemented before?

(Even Oldridge) #2

What you’re talking about is called transfer learning and it’s incredibly popular amongst image processing. It’s only recently been introduced to NLP, but Jeremy was one of the main contributors of the idea there.

But it’s very straightforward as long as you have the data.

(John Wu) #3

Thanks for the reply @Even! It makes sense that this falls under the umbrella of transfer learning, although I was under the mistaken assumption that transfer learning simply meant using a pretrained model for another data set. Now I can proceed with confidence!