4 ms·

Horrifying usage aside... I'm curious about the tech here as it relates to structural bias. In the US there are frequent stories about how ML works well on whi

Horrifying usage aside... I'm curious about the tech here as it relates to structural bias. In the US there are frequent stories about how ML works well on white men, but not, e.g., African Americans and I think Asian Americans. There was that Twitter spat between Timnit Gebru and Andrej Karpathy about this subject, and whether it's "just" training data or inherent in ML, partly because of who runs the analyses and who the wealthy consumers are, etc. Having trained some models and fiddled with hyperparameters myself, I could see the argument that it's a little more than just the training data; for example, how the algorithm uses contrast or what kind of angles it picks up. But what I'm curious about is how well this system works here, developed by a Chinese company with Chinese developers for primarily Chinese users. Does it have better accuracy than Google on the same Chinese data set? Does it have worse accuracy on white people than Google? I imagine the data is not out there, but maybe there's some published research along these lines by Tencent, or others comparing algorithms or models trained on Tencent data sets to the western equivalents?

godelski 5y ago
A lot of that bias comes from the dataset used. CelebA is a common one but not the only one. In these datasets the faces are predominantly white. There's also the difficulty that darker faces are more likely to not gather as much detail given certain lighting conditions and so that affects Black people more even if you introduce an equal number of samples (so you need more diversity in lighting conditions for them). That said, Baidu and Tencent have datasets of predominantly Asian faces, so their bias should go in the other direction. Being better at Asian faces (probably they do well on white faces because they can pull from typical research databases), and then not so well at Black faces. But this is just a guess. We can't know without seeing the results (which then it is still hard to determine bias because of how results are presented).

godelski 5y ago

A lot of that bias comes from the dataset used. CelebA is a common one but not the only one. In these datasets the faces are predominantly white. There's also the difficulty that darker faces are more likely to not gather as much detail given certain lighting conditions and so that affects Black people more even if you introduce an equal number of samples (so you need more diversity in lighting conditions for them). That said, Baidu and Tencent have datasets of predominantly Asian faces, so their bias should go in the other direction. Being better at Asian faces (probably they do well on white faces because they can pull from typical research databases), and then not so well at Black faces. But this is just a guess. We can't know without seeing the results (which then it is still hard to determine bias because of how results are presented).