I am a high school senior about to start my undergraduate education. I've done an internship for the past two years working with constitutional neural networks and RNNs. I love doing research, and I know that's what I want to do in the future, and given my interest and proficiency in the field I had assumed that I would research AI.

Recently, though, I've begun to rethink this. There has been more and more research coming out that points to the automation of AI research (for example, these three papers). Speaking about the third paper, Nando de Freitas joked in his recent talk, "If you're designing gradient descent algorithms for MNIST, you're out of a job."

We are not at the end of machine learning research, but will we be in another 10, 20, 30 years? Will we reach a point in the next few decades where there is no more major AI research to be done by humans? Starting college, I want to go into a field with a rich future and great potential for innovation. Will deep learning still be that in 10 years when I'm looking for a job?

  • $\begingroup$ Well, isn't machines building all the stuff the future? Go from there... Don't take individual comments too seriously but rather look at the overall market situation. Here is a good resource: bls.gov/oes/current/oes_nat.htm#15-0000 $\endgroup$ – Tony Jan 21 '17 at 22:05
  • $\begingroup$ If AI/ML work is automated, it seems plausible that as an AI/ML researcher, you would be doing the work around automating it. Moreover, automatic methods are not able to integrate information about subject matter: working on AI for national security is a different topic than AI for malware analysis, and knowing how to extract features from these different sources of data is a key part of the task. $\endgroup$ – Sycorax Jan 21 '17 at 22:10
  • $\begingroup$ I've a younger brother (18 yrs), and I have pushed on him to become as technical as possible. I'll pass on the same suggestions, whether you end up working with neural nets or some other CS/engineering field, it will benefit you to be a proficient programmer in languages like C++, Python, Java. Also, learn to use a Linux/UNIX shell. $\endgroup$ – Jon Jan 21 '17 at 22:26
  • $\begingroup$ The only thing that's changed in recent years is the speed at which new algorithms are developed. So while individual types of algorithms risk being outdone by better techniques the field is nowhere near saturation in terms of what's left to be discovered. I'll bet money that 10 years from now there will be something out there that blows our current deep learning techniques out of the water, maybe even not using neural networks at all. $\endgroup$ – Alex R. Jan 21 '17 at 22:47

My father went to university shortly after World War II and he used to tell me about how he once did a long Fourier transformation and how they used those very long paper rolls, as they did it in a paper pencil approach.

When I went to school in the 80s, people used to think, that in some time in the future everybody would have to deal with computers. And we supposed to be the generation who could not avoid computers. Not everybody did believe it, but school tried to prepare us by teaching us BASIC and PASCAL!

Now it turns out, we still need a lot of people who know about Fourier transformations and yes, next to everybody uses computers today. But how wrong were they, teaching programming to everyone, because they believed, that "working with computers" meant "programming computers"?

I met a young automobile engineer last year and he told me, how all those engineering students went into "automobile" because they somehow love fuel burning engines and now see, that electric mobility is what they will earn their money with, in the future.

Obviously, doing computer analyses in 10, 20, 30 years will be totally different from what it is today. And nobody can tell you now, what it will be like. There is good reason to conject, that data analysis is going to be of huge importance and that some of the jobs done today by specialists will be done in 20 years by anyone with his watch or mobile or... But new questions will come up, new things to learn and to master, things that will need specialists (at least for a time). Don't start a career in information technology assuming, that you won't have to learn and to adapt to new technologies your whole life through. Try to learn a lot, try to be fit for change. Learn those things, that will never change (mostly mathematics) and don't overdo that fancy new method, that everyone praises now just to reject it, when the next hot thing comes around.

SQL has been used 40 yrs ago and is used today. For someone who knows their linear algebra, their basics in databases, knows about object oriented programming as well as functional programming, classical statistics as well as Baysian and Machine Learning, can communicate with ordinary people, with business people and lawyers and is constantly open to new challenges, there will be lots of interesting work opportunities in the future. We just don't know, what they will be.

So if you love all the consituents of what makes machine learning, go for it. If just think, that support vector machines are cool - well, think twice.


Not the answer you're looking for? Browse other questions tagged or ask your own question.