5 ms·

Only white people used as examples. Might be useful to literally diversify your dataset.

by debt 8y ago

lopmotr 8y ago
They're also all humans. How diverse they go depends on the application people put it to. It's surely enough to prove the concept with one race and one species.
wpietri 8y ago
Absolutely. Even if one doesn't care about quality or inclusion, making this mistake can lead to big PR fails (and apparently lasting tech struggles): https://www.wired.com/story/when-it-comes-to-gorillas-google-photos-remains-blind/ https://www.wired.com/story/when-it-comes-to-gorillas-google...
jahewson 8y ago
This is somewhat ironic given that all five of the paper’s authors are East Asian.
- TaylorAlexander 8y ago
  Fei Fei Li recently pointed out in a talk that if you search google images for “grandma” you get very biased results. They’re all white grandmothers! So it may be that the data sources the authors had are biased too. In any case this is an issue that should be addressed.
  - jurjenh 8y ago
    That made me wonder about what would happen if google started grouping concepts (eg words for grandmother in other languages will almost certainly give you differing results) rather than words... Possibly (probably) the cognitive load on people might lead to bad A/B test results, but it would be a curious thing to explore - cultural bias would get a bit of a beating I'd imagine.
  - kowdermeister 8y ago
    > In any case this is an issue that should be addressed Why is that? This is just a demonstration of a new idea with existing technology, not some ready to deploy app for 2 billion people.
DanAndersen 8y ago
The arxiv paper [0] shows some further examples, which include some non-white faces. Your point is a fair one in the case of this application, since the goal is to do 3D reconstruction of the face. The project's YouTube video [1] shows one example (around 0:44) of the reconstruction run on test input of a black woman, and the reconstruction has some troubles modeling the wider nose common in black populations. More varied data (or an ensemble method that attempts coarse racial categorization and then further refining ethnicity-specific face models) might get better results. [0] https://arxiv.org/pdf/1803.07835.pdf https://arxiv.org/pdf/1803.07835.pdf [1] https://www.youtube.com/watch?v=tXTgLSyIha8&feature=youtu.be https://www.youtube.com/watch?v=tXTgLSyIha8&feature=youtu.be

Only white people used as examples. Might be useful to literally diversify your dataset.

by debt 8y ago

Only white people used as examples. Might be useful to literally diversify your dataset.

They're also all humans. How diverse they go depends on the application people put it to. It's surely enough to prove the concept with one race and one species.

wpietri 8y ago

Absolutely. Even if one doesn't care about quality or inclusion, making this mistake can lead to big PR fails (and apparently lasting tech struggles): https://www.wired.com/story/when-it-comes-to-gorillas-google-photos-remains-blind/ https://www.wired.com/story/when-it-comes-to-gorillas-google...

jahewson 8y ago

This is somewhat ironic given that all five of the paper’s authors are East Asian.

TaylorAlexander 8y ago
Fei Fei Li recently pointed out in a talk that if you search google images for “grandma” you get very biased results. They’re all white grandmothers! So it may be that the data sources the authors had are biased too. In any case this is an issue that should be addressed.
- jurjenh 8y ago
  That made me wonder about what would happen if google started grouping concepts (eg words for grandmother in other languages will almost certainly give you differing results) rather than words... Possibly (probably) the cognitive load on people might lead to bad A/B test results, but it would be a curious thing to explore - cultural bias would get a bit of a beating I'd imagine.
- kowdermeister 8y ago
  > In any case this is an issue that should be addressed Why is that? This is just a demonstration of a new idea with existing technology, not some ready to deploy app for 2 billion people.

DanAndersen 8y ago

The arxiv paper [0] shows some further examples, which include some non-white faces. Your point is a fair one in the case of this application, since the goal is to do 3D reconstruction of the face. The project's YouTube video [1] shows one example (around 0:44) of the reconstruction run on test input of a black woman, and the reconstruction has some troubles modeling the wider nose common in black populations. More varied data (or an ensemble method that attempts coarse racial categorization and then further refining ethnicity-specific face models) might get better results. [0] https://arxiv.org/pdf/1803.07835.pdf https://arxiv.org/pdf/1803.07835.pdf [1] https://www.youtube.com/watch?v=tXTgLSyIha8&feature=youtu.be https://www.youtube.com/watch?v=tXTgLSyIha8&feature=youtu.be