11 ms·

# EDIT: it seems the post below me is actually correct, and my post is incorrect in this case. Just as an aside here: "Blondeness" and "beard" are probably jus

# EDIT: it seems the post below me is actually correct, and my post is incorrect in this case. Just as an aside here: "Blondeness" and "beard" are probably just labels the authors found correspond the most to the latent variables in this case. This means that there won't be a perfect translation between those words and what these variables directly respond to in the network. So although the training data may have been biased with more smiling blonde people, it doesn't necessarily have to have been so. It might be that what this latent variable encodes just does something else in edge cases where there are few examples.

matheist 8y ago
Not so. They split the data between "bearded faces" and "non-bearded faces" and compute the vector from one to the other, and use that to alter a given face. (It reminds me of the typical word2vec example of man - woman + queen = king.) See their code snippet halfway down their page, or "Semantic Manipulation" on page 8 of their linked paper.
- nerdponx 8y ago
  This sounds like yet another appearance of the tank-sky problem.
  - blt 8y ago
    There is no mechanism in this paper (or in a standard VAE, or GAN) to encourage that a single human-understandable semantic quantity should be captured in a single dimension of the latent code. So, in general, it won't happen. "Blondeness" is spread out over all dimensions of the latent code. Therefore the method in this paper (summarized by matheist) is totally reasonable. There are other autoencoder-type generative models that try to concentrate each attribute in one dimension, usually by using the class labels as an additional input. But that is not the focus of this paper.
    - hanrelan 8y ago
      Do you have a link to a paper that does this by any chance? I'm interested in learning more but not sure what to search for.
subcosmos 8y ago
Sounds like the approach used here : https://arxiv.org/abs/1609.04468 https://arxiv.org/abs/1609.04468

matheist 8y ago

Not so. They split the data between "bearded faces" and "non-bearded faces" and compute the vector from one to the other, and use that to alter a given face. (It reminds me of the typical word2vec example of man - woman + queen = king.) See their code snippet halfway down their page, or "Semantic Manipulation" on page 8 of their linked paper.

nerdponx 8y ago
This sounds like yet another appearance of the tank-sky problem.
- blt 8y ago
  There is no mechanism in this paper (or in a standard VAE, or GAN) to encourage that a single human-understandable semantic quantity should be captured in a single dimension of the latent code. So, in general, it won't happen. "Blondeness" is spread out over all dimensions of the latent code. Therefore the method in this paper (summarized by matheist) is totally reasonable. There are other autoencoder-type generative models that try to concentrate each attribute in one dimension, usually by using the class labels as an additional input. But that is not the focus of this paper.
  - hanrelan 8y ago
    Do you have a link to a paper that does this by any chance? I'm interested in learning more but not sure what to search for.

subcosmos 8y ago

Sounds like the approach used here : https://arxiv.org/abs/1609.04468 https://arxiv.org/abs/1609.04468