Thanks for the link. This surprised me: “These high retention rates are important, because a large majority (about 2/3) of grad students getting trained in AI-related programs at US universities are not American citizens”
The current situation of making it difficult for recent graduates to get green cards could put pressure on retention (having students staying in the US). At my last job I tried for about a year to replace a deep learning engineer after someone on my team left for a dream job.
Currently many white collar jobs (like Wall Street analysts) are getting eliminated by deep learning and other models automation. I predict that it will be 5 to 7 years before most routine deep learning development will be largely automated, but until we achieve this, it is really important to have trained people to do the work. That said, top level AI research will not be automated in the near future.
Is any sizable portion of these jobs interchangeable with deep learning?
The best deep learning engineers with infinite hardware built Amazon’s recommendation algorithm, and it’s largely trash. I can’t imagine it’s magically thousands of times more capable for these other tasks.
> built Amazon’s recommendation algorithm, and it’s largely trash
What is trash for you may not actually be reflective of how good the algorithm is. My experience is that Amazon's algorithm actually works somewhat well.
Recommendation algorithms usually work well for people who are closer to average (i.e. their preferences being closer to the aggregate). People who deviate from the average tend to get poorer recommendations. It does even worse for segments that deviate so far from average where there are few sample points.
Recommendation algorithms optimize for aggregate conversion -- if it can get the most number of people to convert, it is good (for the merchant) even if it looks like your personal recommendations suck. It's all a matter of perspective on what "good" is.
And the statement you made above "Currently many white collar jobs (like Wall Street analysts) are getting eliminated by deep learning and other models automation." is complete bullshit. I don't care if your other name is Yann Lecun; that statement is completely false.
This topic has been in the news a lot in 2019. Try a web search for “wall street analysts replaced by ai deep learning”. Again, just my opinion: formerly high paying jobs will be lost, with the best people in each category retained and their work augmented with next generation AI and otherwise automated systems.
Look, there are a ton of marketing submarines[1] on this topic, and there is precious little fact checking involved. I've worked on numerous wall street projects, I presently advise several HF startups; higher frequency and low, have consulted for a couple of fintech companies in the recent past, I've worked with advanced research teams at JPMC, Citi, Credit Suisse, a couple of big HF you've heard of and smaller ones you've never heard of. I regularly talk to one of the best regarded academic guys who attempts to apply DL to trading problems. I keep my ear to the ground as to the latest trends; people on the street talk to me because I'm a funny asshole and I know what I'm talking about. I have found zero evidence DL is being used in the financial services industry at all.
The reasons for this should be quite clear. It isn't appropriate for this kind of work. Pretty much all financial problems are insanely noisy, and DL is shit at dealing with any kind of noise, both in labeled, unlabeled and reinforcement configurations. Trading problems on short time horizons can harvest innovations vastly faster than DL can keep up using what amounts to Kalman filter like ideas. Long time horizons don't have enough data to fit a meaningful DL model even if 90% of the spectrum weren't noise. We're not talking something like automatic translation where you basically have all the clean data on the internet to fit to. The only place I could imagine it being used at the two-sigmas of the world is in natural language processing, but since you are asserting something completely different, that's irrelevant.
R&D departments, of course, fool around with it. And fintech companies are definitely implying they're using it because ... everyone else is. IRL they're all doing some kind of Markowitz MPT with factor models spooge. If you have any actual evidence that there is some large scale automation of wall street analysts facilitated by deep learning: I'd love to see it. Marketing submarines riding the DL hype chuckwagon are not evidence.
The current situation of making it difficult for recent graduates to get green cards could put pressure on retention (having students staying in the US). At my last job I tried for about a year to replace a deep learning engineer after someone on my team left for a dream job.
Currently many white collar jobs (like Wall Street analysts) are getting eliminated by deep learning and other models automation. I predict that it will be 5 to 7 years before most routine deep learning development will be largely automated, but until we achieve this, it is really important to have trained people to do the work. That said, top level AI research will not be automated in the near future.