Good points.

1. I only accepted images that yielded 50% confidence on my existing model and then manually sorted out the rest. As the model stands, based on two additional labeled datasets, >50% confidence had accuracy over 98% on https://data.vision.ee.ethz.ch/cvl/rrothe/imdb-wiki/
2. I did it here (https://medium.com/@synopsi/thats-very-cool-f2aa83d2a73e) and you can see the results are weaker (more manual work for me).
3. Don’t believe so. But I would have to have much larger, much cleaner and much broader dataset.
4. It actually does make sense as it somehow correlates with the years people change jobs in Silicon Valley (much more often than in other parts of the country). Also you can see it’s more extreme for women as it may correlate with the time they have children.

But overall the dataset is extremely limited so it will not produce much better data. I wasn’t trying to solve any significant problems just answer my own questions. If nothing else, it told me that I have to pay better attention when hiring women to not discriminate them.