Summary of the Neural Information Processing Systems Conference 2016
blog.evjang.com
blog.evjang.com
Maybe Apple, Facebook, and Google think they can develop their talent in house from their programmers?
http://ijcai-16.org/index.php/welcome/view/sponsors
And at ICML
http://icml.cc/2016/?page_id=63
Have been to Google booths at IJCAI and got selected for interviewing based on my interaction at the booths.
To whatever extent they have trouble attracting researchers it's because you will never be able to publish anything or talk about your work in a finance firm. Academia is status/reputation/ego driven (not that there's anything wrong with this).
My friends in quant finance made more than 200k guaranteed their first years (not including performance bonus), and they did not have graduate degrees. It went up to more than half a million in a few years (including bonus).
The ramp up and ceiling is much higher in finance. First year salary for someone with a bachelors in a finance company may be comparable to someone with a masters/PhD in a tech company, sure.
The common complaint is that regardless of performance, their pay increases and bonuses end up being shit YOY and that the only way to make serious money is to hop from company to company every year.
One friend who builds trading infrastructure has increased his pay from $180k to north of $480k just by doing this and is about to go to his 5th or 6th fund (in roughly as many years). He kills it, ends up getting the token bonus they give tech people, waits out his non-compete and leaves.
They treat tech talent as commodity just as much as anywhere else. The other downsides being you usually can't trade your own account while employed.
Said friend would prefer to start a fund soon rather than keep this up as he'd rather be able to trade.
Publishing isn't the focus of hedge funds so they don't get a lot of love.
It makes sense that academics would be interested in working for a company that will still allow the academic to publish.
He makes a great point I think inadvertently:
However, they haven't created significant commercial value in industry yet, in ways that couldn't plausibly be substituted with traditional supervised learning.
Transfer learning, domain adaptation and semi-supervised learning alleviate the data-hungry requirements of deep learning, and are starting to work really well.
Transfer learning is a term of art, but it also represents how supervised learning works from human>machine.
I think we need to embed more ML systems into our daily lives to "teach" networks how/when/why to do things. IMO the best way to do this is through AR as it's a great input output tool for recommending actions to the user and transmitting the type of data that we are making great progress in (vision).
A part of me feels very scared. What if the big companies do end up replacing millions of jobs.
Better than human driving, speech recognition, image recognition, translation, learning to code, knowledge systems.
Once you learn how to create a better human brain and make billions of them at scale. The big companies become monoliths who can get as many AI slaves as they need.
Do we even need 10 billion people anymore?
"many top hedge funds and trading shops came to NIPS to run career booths, but there was a surprising lack of interest from attendees (attendees were more interested in the likes of Apple, Facebook, Deepmind, Google, etc).
At a regular college career fair in the East Coast, these roles are reversed. The talent pool at NIPS seems to be far more interested in tech and open-ended research than making money."