2

For instance, if I wanted a train network that can output

  • van
  • truck
  • sedan
  • vehicle
  • pedestrian

does it make sense to only train it on van, truck, sedan, and pedestrian and then make "vehicle" a synonym of van, truck, sedan? Or does it make sense to train vehicle as its own class made up of the combined training data of van, truck, sedan?

EDIT ---

Here's a more clear example I think, what if there are labels that are not exclusive? For instance, a child can also be a pedestrian. How do I train a network to output both labels if it, for example, sees a child walking down the street? It seems that using a softmax later as the final layer of the network would not allow this to work as all probabilities need to sum to one, in this case it could be 90% child and 80% pedestrian and that's entirely valid.

John Allard
  • 131
  • 4
  • Obviously they all are vehicles... Can't see the reason of the question? – Aditya Aug 28 '18 at 02:41
  • But one of the classes is human, which is not a vehicle. It's like this.. if I want my network to recognize many things, including dogs, cats, humans, and mammals, does it make sense to make mammals a derived label from (dog OR cat OR human) or does it make sense to combine all of the mammal training data (all cats, humans, dogs) into a single class, mammal, and feed that into the network in addition to training it on the specific mammal classes? – John Allard Aug 28 '18 at 20:19
  • Just append an 'if' statement in your code. – Daniel Aug 28 '18 at 21:07
  • Also, look at multimodal classification – Daniel Aug 28 '18 at 21:08
  • It seems like your edit totally changed the question, and the new one should be asked separately as its own question. In the future, can you take care to ensure that the question accurately represents what you want answered, so people don't spend time answering something that isn't what you want answered? Thank you! – D.W. Aug 28 '18 at 21:16
  • You're right, I was not very clear and that was not fair of me to change up the question like that with an edit. I have thought through the problem and have made a new question that I hope is more clear. That question is here: https://datascience.stackexchange.com/questions/37542/how-to-train-dnn-to-output-labels-that-are-not-exclusive-of-one-another – John Allard Aug 28 '18 at 23:01

1 Answers1

1

It depends what you're trying to achieve.

In general, there are things that are vehicles but aren't vans, trucks, or sedans, so your two proposed schemes are not equivalent even if you were somehow able to obtain a 100% accurate classifier.

In general, the way to answer questions like this is to try it both ways and see which works better.

D.W.
  • 3,341
  • 17
  • 42