> “We risk losing essential information,” she says. “The datasets need to reflect the real statistics in the world.”
I think it comes down to that. If we feed a machine a gross of images of women shooting, a gross of images of women feeding babies, a gross of images of men shooting and a gross of images of men feeding babies (just to pick two stereotypical activities at random — in my family men & women both shoot and take care of kids), then I suspect it'll generate worse results than if we feed it a set of images reflecting reality.
> Yatskar describes a future robot that when unsure of what someone is doing in the kitchen offers a man a beer and a woman help washing dishes. "A system that takes action that can be clearly attributed to gender bias cannot effectively function with people," he says.
I dunno, that sounds like it'd effectively work with people. The rule 'if unsure then if isMan then offer beer else offer help with dishes' sounds like a reasonable heuristic for someone learning social skills. Indeed, it sounds like something a kid would come up with.
This all goes back to the utility of stereotypes. Stereotypes are just patterns; if a human being or a learning machine identifies true patterns and uses them, then he, she or it will get better results then otherwise.
Stereotypes are not "just patterns". Google "stereotype" and the first definition is "a widely held but fixed and oversimplified image or idea of a particular type of person or thing.".
In this case, the machine is learning, for example, that women do dishes and men drink beer. This isn't based on empirical data and patterns. It comes from the data the algorithm is trained on, which in this case is "...more than 100,000 images of complex scenes drawn from the web, labeled with descriptions". Those descriptions inevitably reflect human stereotypes (which again aren't just patterns). "Both datasets contain many more images of men than women, and the objects and activities depicted with different genders show what the researchers call “significant” gender bias."
Should machines understand that men are more likely to be construction workers than women, I think so. But that doesn't mean that biased data is not a huge problem.
At the very least, we should strive to teach machines to understand the world as it is, not as we view it through flawed, biased eyes. Datasets generated by humans are going to be invariably flawed and far from reality, unless we take careful steps to ensure otherwise. Emphasizing the importance of these datasets, we should be particularly concerned when we are teaching machines to act based on our biases and find that (in the case of this article), the machine is actually learning to amplify our own biases.
Patterns do fit stereotypes perfectly. They are oversimplified concepts of something real; some samples are outside the pattern and that is perfectly fine for pattern-discovery. So stereotypes and patterns aren't distinct concepts in essence.
«Datasets generated by humans are going to be invariably flawed and far from reality, unless we take careful steps to ensure otherwise.»
If the data in the wild is not real, we can only adapt it to your reality to make it real. That's not necessarily my view of reality. You can't just pluck objectivy out of the air and fresh up fake real data.
No they don't. "Stereotype" is a psychological concept, and therefore by definition incorporates human subjectivity. There are various conflicting definitions, but most include the possibility or likelihood that many overstate or even completely falsely construct generalizations.
Or, as Wikipedia[1] states,
> By the mid-1950s, Gordon Allport wrote that, "It is possible for a stereotype to grow in defiance of all evidence."
> Research on the role of illusory correlations in the formation of stereotypes suggests that stereotypes can develop because of incorrect inferences about the relationship between two events (e.g., membership in a social group and bad or good attributes). This means that at least some stereotypes are inaccurate.
I thought the issue was that the machine is inaccurate (although I understand there was always a certain amount of accuracy with ML)
Wasnt the real issue:
"The researchers’ paper includes a photo of a man at a stove labeled “woman.” as stated in the article?
Thus isnt this article rather calling for an improvement in the system and simply stating that the data coming in is not as good because systems are also heavily influenced by data or as they say garbage data generates garbage results.
I think it comes down to people wanting AI to do what we say ought to happen, rather than what we actually do. I wish the solution was to fix what we actually do, rather than have to rejigger the data, but here we are.
All your examples are characteristics, not preferences. The model gets the data 'there aren't many women in STEM (end of pipeline)' so it doesn't allocate resources to girls who are interested at the start of the pipeline. The prediction -> action cycle generates a self fulfilling prediction that winds up at a bad local minimum (maybe the model needed better leakyREU functions).
This is the same problem why your netflix recommendations just show you either the stuff everyone else is watching or the stuff super similar to the stuff you've already liked, it's hard to make predictions that are both novel and useful. There has to be room for exploring new domain space.
This has been known for many years. Women engineers at Google were particularly disappointed when for a lot af them the ads demographic inference algorithms would predict that they're men, based on their interest in engineering.
I think it comes down to that. If we feed a machine a gross of images of women shooting, a gross of images of women feeding babies, a gross of images of men shooting and a gross of images of men feeding babies (just to pick two stereotypical activities at random — in my family men & women both shoot and take care of kids), then I suspect it'll generate worse results than if we feed it a set of images reflecting reality.
> Yatskar describes a future robot that when unsure of what someone is doing in the kitchen offers a man a beer and a woman help washing dishes. "A system that takes action that can be clearly attributed to gender bias cannot effectively function with people," he says.
I dunno, that sounds like it'd effectively work with people. The rule 'if unsure then if isMan then offer beer else offer help with dishes' sounds like a reasonable heuristic for someone learning social skills. Indeed, it sounds like something a kid would come up with.
This all goes back to the utility of stereotypes. Stereotypes are just patterns; if a human being or a learning machine identifies true patterns and uses them, then he, she or it will get better results then otherwise.