-5

I’m reading this paper about sound separation by using neural network method. It uses the term "annealing the temperature parameter in a softmax activation function". What does it mean?

Itamar Mushkin
  • 2,803
  • 2
  • 16
  • 32
user11190123
  • 1
  • 2
  • 1
    Could you post the paper or give reference to what you are referring to? – Chris Farr May 28 '19 at 13:09
  • @ChrisFarr https://www.groundai.com/project/self-supervised-audio-visual-co-segmentation/ Here you can access the paper the phrase I have problem in, is on page two before the first function. – user11190123 May 28 '19 at 13:13
  • Great, you may want to edit your original question with that information as well. – Chris Farr May 28 '19 at 13:20
  • 1
    Also, this question belongs in https://stats.stackexchange.com/. I gave an answer, but in the future, know that stack overflow is for programming, not for the mathematics/statistics behind it... – Itamar Mushkin May 28 '19 at 13:24

1 Answers1

0

"Annealing", in this sense, is an analogy to a chemical process in which temperature is reduced.
Here, too, the 'temperature' term in the softmax is reduced, changing it gradually from a 'soft' sigmoid function to a 'sharp' argmax
Try plotting a softmax function for yourself with different temperatures, you'll see the difference...

Itamar Mushkin
  • 2,803
  • 2
  • 16
  • 32