9. Large Multi-Modal Models (LMMMs) to Incorporate Images, Art & Video - [E2] Wang, Yilun, and Michal Kosinski.

[lkcao]

Post questions here for this week's exemplary readings:

2. Wang, Yilun, and Michal Kosinski. 2018. “Deep neural networks are more accurate than humans at detecting sexual orientation from facial images.” Journal of Personality and Social Psychology 114(2): 246.

[bucketteOfIvy]

This is an interesting paper, but I think it reveals an underlying concern that we might have when working with multimedia content. Specifically, I think the usage of images of people's faces to detect private information highlights that there is a level of intimacy involved with the usage of image data that potentially goes beyond that seen with text data. How should we best balance ethical these concerns with the need to use large amounts of image data to train models or answer questions?

[yunfeiavawang]

This paper is an excellent example of applying deep neural networks to capture facial features. This is a controversial one, I have to admit, ethically. But if we take the perspective of consequentialism, it is necessary to clarify that this study is valuable in addressing the biological mechanism of sexual orientation, and did ignite the public's privacy concerns at the time it was published. No serious harm has been found in any of the participants up to now. More importantly, media coverage of the critiques on this study attracted copious attention to the growing privacy risk in the digitalized life, reminding online users to take action to protect their personal information and forcing the media platforms to make reasonable rules on protecting users' privacy. I am curious about what steps did the companies or government take after the publication of this paper. Also, is there any change in citizen's perception of image data privacy over the last ten years?

[volt-1]

Given the societal context at the time, I believe the social warning significance of this article above its technological importance. When ml-driven decisions may exacerbate social biases, what moral responsibility should developers take?

[ethanjkoz]

A key point of this article are the potential technological and ethical ramifications for a model that can predict sexual orientation with just one's face. The authors relate prenatal hormone theory as a possible explanation for atypicality of gay men and women's faces. Though the authors certainly understand the potential for technology, I am curious to see how well this approach could be used for extensions that they mention. Are there physical features that reveal political affiliation, as the authors note?

[naivetoad]

How do cosmetic procedures, makeup, or changes in grooming habits affect the DNN's ability to predict sexual orientation from facial images?

[beilrz]

I think this a very interesting paper that examine a common belief, that sexual orientation can indeed by detected through facial feature. I recall reading it last year. I agree there are ethical concern regarding this paper: however, I do feel that ethical problems regarding deep neural net exist long before this paper, while this paper only brought such issue to the public.