|
Hi can some one explain what is Hamming Loss? and what does it mean in context of structured prediction? |
|
Hamming loss is the loss associated with the Hamming distance. The latter is a distance on binary vectors that counts the number of entries that do not agree. As a loss applied to binary vectors this basically counts "how many decisions did you get wrong". See my answer on this question for why that is helpful. To summarize briefly: If I have a sequence labeling problem, like chunking text, it can be decomposed into a BIO encoding. It is difficult to directly optimize the structured prediction loss, e.g. F-measure over chunks. But I can easily optimize the accuracy of the classifier correctly predicting the label for the token.
(Jan 26 at 04:12)
Joseph Turian ♦♦
As a follow-up, is there such a thing as a regularizer that attempts to minimize the hamming distance between similar samples?
(Feb 22 at 23:16)
Brian Vandenberg
Can you explain you setup a bit more?
(Feb 23 at 02:24)
Andreas Mueller
Well, it isn't something I'm working on, the question just occurred to me the other day. The simplest case I can think of is semantic hashing. The papers I've seen on it clearly demonstrate that it works well in principle. But, suppose you wanted to actively force it to hash similar samples to latent representations with the smallest hamming distance. I'm sure other examples could be concocted, that's just roughly what occurred to me.
(Feb 23 at 11:08)
Brian Vandenberg
|
