3 ms·

The fact that some very individually distinctive features are still present on some pictures (why do some of these faces smile, and some don't?), I would guess

The fact that some very individually distinctive features are still present on some pictures (why do some of these faces smile, and some don't?), I would guess that the number of pictures that was averaged was ridiculously low, 10 maybe. The result is therefore all about the original ~10 pictures that were chosen... and it is certainly not an "average face" (which would require a statistically representative sample, north of 10k for most countries). Chose 10 other people from the same ethnicity and you get a different face.

fleitz 13y ago
This is incorrect, a sample of 10 is statistically representative but has a high margin of error. If you can live with a high MOE then you can use very small samples, often small studies of 10 are used in medicine to see whether further studies should occur. Upon doing 5 minutes of research the data appears to come from FaceResearch which has many thousands of images on file, seems like any issues would be related to sample biases and not sample size.
- azakai 13y ago
  10 is ok for measuring a single variable (although even then, you probably want something like 30, even in the social sciences). But for faces, it is possible that there are large numbers of relevant variables (various features of the eyes, nose, etc. can all vary separately). That increases the sample size you need to get decent results.
  - fleitz 13y ago
    You're allowed to reuse the same sample for multiple variables, pollsters do it all the time, they'll phone one person and ask them 10 different questions. The fact that you're measuring faces as opposed to political leanings does not affect the underlying statistical methods. Confounding variables only come into play when you're doing regressions, and/or trying to prove causation. Since they aren't attempting to explain why the faces look like this there aren't any confounding variables.
    - graycat 13y ago
      Thanks for raising the level of discourse by several notches!
    - azakai 13y ago
      That pollsters do something does not make it valid! ;) Yes, they do ask dozens of questions of a small number of people. But this is also mathematically dubious. If you do multiple random samples, the chance at least one is way off increases. If you have a 95% chance of being close to correct on your small sample, and you do 20 samples, it is quite likely you are wrong on at least 1 of them. My claim is that (1) there is a very large number of relevant facial features for the goal of the survey (find the "average" face), and (2) that being wrong on even a small number of them is significant. My basis for both is a combination of intuition and that we know humans are extremely good at perceiving faces and subtle facial differences. There are parts of the brain that are apparently adapted specifically to facial recognition, and disorders where facial recognition is impaired but nothing else, for example.

fleitz 13y ago

This is incorrect, a sample of 10 is statistically representative but has a high margin of error. If you can live with a high MOE then you can use very small samples, often small studies of 10 are used in medicine to see whether further studies should occur. Upon doing 5 minutes of research the data appears to come from FaceResearch which has many thousands of images on file, seems like any issues would be related to sample biases and not sample size.

azakai 13y ago
10 is ok for measuring a single variable (although even then, you probably want something like 30, even in the social sciences). But for faces, it is possible that there are large numbers of relevant variables (various features of the eyes, nose, etc. can all vary separately). That increases the sample size you need to get decent results.
- fleitz 13y ago
  You're allowed to reuse the same sample for multiple variables, pollsters do it all the time, they'll phone one person and ask them 10 different questions. The fact that you're measuring faces as opposed to political leanings does not affect the underlying statistical methods. Confounding variables only come into play when you're doing regressions, and/or trying to prove causation. Since they aren't attempting to explain why the faces look like this there aren't any confounding variables.
  - graycat 13y ago
    Thanks for raising the level of discourse by several notches!
  - azakai 13y ago
    That pollsters do something does not make it valid! ;) Yes, they do ask dozens of questions of a small number of people. But this is also mathematically dubious. If you do multiple random samples, the chance at least one is way off increases. If you have a 95% chance of being close to correct on your small sample, and you do 20 samples, it is quite likely you are wrong on at least 1 of them. My claim is that (1) there is a very large number of relevant facial features for the goal of the survey (find the "average" face), and (2) that being wrong on even a small number of them is significant. My basis for both is a combination of intuition and that we know humans are extremely good at perceiving faces and subtle facial differences. There are parts of the brain that are apparently adapted specifically to facial recognition, and disorders where facial recognition is impaired but nothing else, for example.