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
This raises some really interesting questions.

We certainly don't want to perpetuate harmful stereotypes. But is it a flaw that the model encodes the world as it really is, statistically, rather than as we would like it to be? By this I mean that there are more light-skinned people in the west than dark, and there are more women nurses than men, which is reflected in the model's training data. If the model only generates images of female nurses, is that a problem to fix, or a correct assessment of the data?

If some particular demographic shows up in 51% of the data but 100% of the model's output shows that one demographic, that does seem like a statistics problem that the model could correct by just picking less likely "next token" predictions.

Also, is it wrong to have localized models? For example, should a model for use in Japan conform to the demographics of Japan, or to that of the world?

>>tines+E7
It depends on whether you'd like the model to learn casual or correlative relationships.

If you want the model to understand what a "nurse" actually is, then it shouldn't be associated with female.

If you want the model to understand how the word "nurse" is usually used, without regard for what a "nurse" actually is, then associating it with female is fine.

The issue with a correlative model is that it can easily be self-reinforcing.

>>karpie+m9
Additionally, if you optimize for most-likely-as-best, you will end up with the stereotypical result 100% of the time, instead of in proportional frequency to the statistics.

Put another way, when we ask for an output optimized for "nursiness", is that not a request for some ur stereotypical nurse?

>>jdashg+9b
You could simply encode a score for how well the output matches the input. If 25% of trees in summer are brown, perhaps the output should also have 25% brown. The model scores itself on frequencies as well as correctness.

>>jvalen+Oc
The only reason these models work is that we don’t interfere with them like that.

Your description is closer to how the open source CLIP+GAN models did it - if you ask for “tree” it starts growing the picture towards treeness until it’s all averagely tree-y rather than being “a picture of a single tree”.

It would be nice if asking for N samples got a diversity of traits you didn’t explicitly ask for. OpenAI seems to solve this by not letting you see it generate humans at all…