I looked at this paper the other day and it looked to me like the non-criminal faces examples were men with shirt collars, whereas the criminal examples were men in t-shirts. If they got above random accuracy, then I wonder if they simply overfitted on that, and the lighting and colour differences between the two styles of photos.
Skyscanner is based in Edinburgh, with another large office in Glasgow. This is huge news for the Scottish tech scene, and will hopefully bring much needed cash to help it thrive. There's a lot of interesting stuff happening here in all types of tech.
Sadly Scotland's other unicorn, FanDuel, looks set to close its hq in Edinburgh after merging with draft kings. However, that will probably release a lot of talent to help smaller co's
To avoid the LinkedIn issue, I crawled job ads (about 10M uk ads over two years), and applied analysis to them. I love your approach - I'm going to have a play this weekend :)
LinkedIn has so many amazing datasets locked up in the platform. I was keen to see whether I could reproduce the skill graph that LinkedIn has, and apply it to other problems, such as valuing and comparing skill sets and knowledge from job titles alone.
This is an amazing dataset. I did some work with it in 2012 at a hackathon, in particular linking it to wiki entities. I used the redirects as drug synonyms and automatically extracted and linked drug names to their Wikipedia pages, which then offered more structured info on the drugs (eg side effects etc). From that, I then did some basic NLP to identify diseases that mentioned those drugs, thereby linking the diseases each drugs can be used to treat for further analysis. The next step we didn't do much with was linking it to the British national formulary - the uk clinical guidelines on drug use (a digital version wasn't readily available then). You can see some of those wiki datasets on my website (http://www.stewh.com - I'm on a train so can't get the exact url, but you should find it in the menu). I had more datasets and the MapReduce code for working with the dataset and wiki dumps if anyone could use it.