zlacker

>>kevema+(OP)
>While we leave an in-depth empirical analysis of social and cultural biases to future work, our small scale internal assessments reveal several limitations that guide our decision not to release our model at this time.

Some of the reasoning:

>Preliminary assessment also suggests Imagen encodes several social biases and stereotypes, including an overall bias towards generating images of people with lighter skin tones and a tendency for images portraying different professions to align with Western gender stereotypes. Finally, even when we focus generations away from people, our preliminary analysis indicates Imagen encodes a range of social and cultural biases when generating images of activities, events, and objects. We aim to make progress on several of these open challenges and limitations in future work.

Really sad that breakthrough technologies are going to be withheld due to our inability to cope with the results.

>>daenz+b5
The big labs have become very sensitive with large model releases. It's too easy to make them generate bad PR, to the point of not releasing almost any of them. Flamingo was also a pretty great vison-language model that wasn't released, not even in a demo. PaLM is supposedly better than GPT-3 but closed off. It will probably take a year for open source models to appear.

>>visarg+I6
That's because we're still bad about long-tailed data and that people outside the research don't realize that we're first prioritizing realistic images before we deal with long-tailed data (which is going to be the more generic form of bias). To be honest, it is a bit silly to focus on long-tailed data when results aren't great. That's why we see the constant pattern of getting good on a dataset and then focusing on the bias in that dataset.

I mean a good example of this is the Pulse[0][1] paper. You may remember it as the white Obama. This became a huge debate and it was pretty easily shown that the largest factor was the dataset bias. This outrage did lead to fixing FFHQ but it also sparked a huge debate with LeCun (data centric bias) and Timnit (model centric bias) at the center. Though Pulse is still remembered for this bias, not for how they responded to it. I should also note that there is human bias in this case as we have a priori knowledge of what the upsampled image should look like (humans are pretty good at this when the small image is already recognizable but this is a difficult metric to mathematically calculate).

It is fairly easy to find adversarial examples, where generative models produce biased results. It is FAR harder to fix these. Since this is known by the community but not by the public (and some community members focus on finding these holes but not fixing them) it creates outrage. Probably best for them to limit their release.

[0] https://arxiv.org/abs/2003.03808

[1] https://cdn.vox-cdn.com/thumbor/MXX-mZqWLQZW8Fdx1ilcFEHR8Wk=...