13

I want to know what it is, and how it is any different from ensembling?

Suppose, I want to achieve high accuracy in classification and segmentation, for a specific task, if I use different networks, such as CNN, RNN, etc to achieve this, is this called an end to end model? (architecture?) or not?

amoeba
  • 93,463
  • 28
  • 275
  • 317
Hossein
  • 2,005
  • 3
  • 18
  • 32

1 Answers1

12
  • end-to-end = all parameters are trained jointly (vs. step-by-step)
  • ensembling = several classifiers are trained independently, each classifier makes a prediction, and all predictions are combined into one using some strategy (e.g., take the most common prediction across all classifiers).
Franck Dernoncourt
  • 42,093
  • 30
  • 155
  • 271
  • Thank you very much. basically I want to achieve a high accuracy in classification and segmentation, and I am willing to use deep learning methods (architectures?) and form one complete model out of all of these methods and architectures. Is this called end to end? is is it considered ensembling? – Hossein Jul 16 '16 at 18:44
  • @Hossein are your classifiers trained independently? – Franck Dernoncourt Jul 16 '16 at 18:45
  • Thats a nice question, I have no idea! How does that affect us? you know I'm researching at the moment, and I have no idea what I'm going to be dealing with. So I am trying to get a good grasp on the concepts I see in literature. – Hossein Jul 16 '16 at 18:48
  • @Hossein it only affects the name :) – Franck Dernoncourt Jul 16 '16 at 18:52
  • Lets put it this way, I have a classifier, probably a CNN, and then I need a segmentation part, this may be an RNN, so basically what I want is a pipe line which does classification and segmentation (like FCN?!) , so I myself think this cant be called an enssemble, since there is probably one classifier, which is accompanied with other architectures and algorithms to achieve one goal. – Hossein Jul 16 '16 at 18:56
  • 1
    @Hossein: If you train the CNN at the same time as the RNN, then it's end-to-end training. – Franck Dernoncourt Jul 16 '16 at 19:03