3 ms·

> If the goal is to personalize attractive faces, you could just have subjects rate pictures on a scale of 1-100. However, research in the field (the field her

by jmmcd 6y ago

> If the goal is to personalize attractive faces, you could just have subjects rate pictures on a scale of 1-100. However, research in the field (the field here is really "interactive genetic algorithms") suggests that it's more efficient/better HCI to allow selection rather than numerical rating, in order to choose the "parents" of the next generation. So, the right way to do this would be a typical interactive GA, using a mouse. Or eye-tracking, for lazy users.

klmadfejno 6y ago
In what way is this a genetic algorithm? When I think genetic algorithms, I imagine biologically inspired non-linear optimization algorithms. That feels like its possibly relevant, but more of a choice of technique than the main idea at play here. Am I wrong?
- jmmcd 6y ago
  Well, I think they are using a GA, based on: > Machine learning-based algorithms subsequently determined which faces produced the greatest amount of activity for each person, then established which traits those faces had in common. Based on that data, the neural network then proceeded to produce new faces that combined those traits. In particular, it is a reasonably common idea (in the last few years) to use a GA or other metaheuristic to traverse an embedding space. Here, that embedding space is the z layer of the GAN. Especially when "fitness" in the embedding space is purely subjective, it's natural to use an interactive GA. However, I am over-stating it to say that the field is just interactive GA. It's that, intersected with face-GANs.
  - klmadfejno 6y ago
    Wouldn't it just be finding the line of best fit in the embedding space, mapping the embedding of each face to the amount of activity?
    - jmmcd 6y ago
      A line of best fit wouldn't quite make sense to me, as the implication would be that extrapolation on that line would lead to improvements -- forever. A GA allows for iterative generation of multiple faces (no linear assumption) and selection of the best. I don't have access to the article so can't check what they actually did!

by jmmcd 6y ago

klmadfejno 6y ago

In what way is this a genetic algorithm? When I think genetic algorithms, I imagine biologically inspired non-linear optimization algorithms. That feels like its possibly relevant, but more of a choice of technique than the main idea at play here. Am I wrong?

jmmcd 6y ago
Well, I think they are using a GA, based on: > Machine learning-based algorithms subsequently determined which faces produced the greatest amount of activity for each person, then established which traits those faces had in common. Based on that data, the neural network then proceeded to produce new faces that combined those traits. In particular, it is a reasonably common idea (in the last few years) to use a GA or other metaheuristic to traverse an embedding space. Here, that embedding space is the z layer of the GAN. Especially when "fitness" in the embedding space is purely subjective, it's natural to use an interactive GA. However, I am over-stating it to say that the field is just interactive GA. It's that, intersected with face-GANs.
- klmadfejno 6y ago
  Wouldn't it just be finding the line of best fit in the embedding space, mapping the embedding of each face to the amount of activity?
  - jmmcd 6y ago
    A line of best fit wouldn't quite make sense to me, as the implication would be that extrapolation on that line would lead to improvements -- forever. A GA allows for iterative generation of multiple faces (no linear assumption) and selection of the best. I don't have access to the article so can't check what they actually did!