Prof. Anima Anandkumar

There's also a fair amount of evidence that women POC aren't well-represented in the training data for *human* doctors :-(

True..but they are not black boxes. I can talk to them and ask questions

This is an interesting point. While I totally dislike the fact that south-asians are largely absent from most major genomic studies. It is also somewhat well known that when you have small sample sized datasets, it's better to have it be as homogenous as possible.

Since small datasets are quite common in clinical studies, it puts me in a very weird position. The researcher side of me hopes the dataset is mostly mono-ethnic (not a technical term) but the human side of me hopes the dataset is diverse so that all groups are represented.

Honestly I am equally surprised by the comment/question. Of course Hinton understands the limitations of the technology but many of his readers, for instance doctors, don’t.

Bringing sexism into this discussion? But.. Why? No real arguments and in need of an easy prop? And picking on the choice of accuracy as the metric? That's seriously missing the point of the argument and it's quite a disappointment.

The fact that women are underrepresented in clinical trials, and generally in medically related datasets is well known. See, e.g. https://www.brighamandwomens.org/assets/bwh/womens-health/pdfs/connorsreportfinal.pdf … @AnimaAnandkumar makes an excellent point.