Ask Slashdot: What Types of Jobs Are Opening Up In the New Field of AI?

Qbertino writes: I'm about to move on in my career after having a "short rethink and regroup break" and was for quite some time now thinking about getting into perhaps a new programming language and technology, like NodeJS or Java/Kotlin or something. But I have the seriously growing suspicion that artificial intelligence is coming for us programmers and IT experts faster than we might want to admit. Just last weekend I heard myself saying to a friend who was a pioneer on the web, "AI is today what the web was in 1993" -- I think that to be very true. So just 20 minutes ago I started thinking and wondering about what types of jobs there are in AI. Is anything popping up in the industry from the AI hype and what are these positions called, what do they precisely do and what are the skills needed to do them? I suspect something like an "AI Architect" for planning AI setups and clearly defining the boundaries of what the AI is supposed to do and explore. Then I presume the requirements for something like an "AI Maintainer" and/or "AI Trainer," which would probably resemble something like an admin of a big data storage, looking at statistics and making educated decisions on which "AI Training Paths" the AI should continue to explore to gain the skill required and deciding when the "AI" is ready to be let go on to the task. You're seeing we -- AFAIK -- don't even have names for these positions yet, but I suspect, just as in the internet/web boom 20 years ago, that is about to change *very* fast.

And what about Tensor Flow? Should I toy around with it or are we past that stage already and will others do AI setup and installation better than me before I know how this thing really works? Because I also suspect most of the AI work for humans will closely be tied to services and providers such as Google. You know, renting "AI" as you rent webspace or subscribe to bandwidth today. Any services and industry vendors I should look into -- besides the obvious Google that is? In a nutshell, what work is there in the field of AI that can be done and how do I move into that? Like now. And what should I maybe get a degree in if I want to be on top of this AI thing? And how would you go about gaining skill and knowledge on AI today, and I mean literally, today. I know, tons of questions but insightful advice is requested from an educated slashdot crowd. And I bet I'm not the only one interested in this topic. Thanks.

Comments

Dowsing

[tietokone-olmi]

But you gotta bring your own "battle-proven" rod.

Difficult

[jawtheshark]

Doing AI is much harder than being an application developer. I doubt most of us would be able to switch. Good luck.

Re:Difficult

[ShanghaiBill]

Doing AI is much harder than being an application developer.

Indeed. I know plenty of companies with openings, but they are looking for PhDs in machine learning and data science from top tier universities. Designing and training a deep ANN is a lot harder than learning how to edit HTML. It is not something you are going to learn in a 21 day "bootcamp".

Re:There is no 'AI'

[Anonymous Coward]

Like 'self-driving' cars AI is a scam and my eyes glaze over when I hear anyone mentioning it. It's usually a clueless investor or someone with something to peddle. All we have are systems with an ever increasing number of if statements. There is no true AI and won't be for a very long time.

If you think self-driving cars are a scam, then yes, AI is definitely a scam. Maybe you should wait a year or so and re-evaluate.
You can't ignore that computers can now beat us at chess, Go, or pretty much any game based on strategy.
Applications such as face-recognition, speech-recognition are already used in every-day life, and are increasingly based on deep-nets.
I'm curious to know how many if statements are in the code?

AI / ML is not a hobby

[jmcbain]

for quite some time now thinking about getting into perhaps a new programming language and technology, like NodeJS or Java/Kotlin or something

You are in for disappointment if you think that getting into a career in AI / ML is anything like a programming hobby, such as picking up NodeJS or Java over a few weekends. First, let me make clear the terminology I'm using. Artificial Intelligence is a broad field. Although the public perception of AI is of software/robots like HAL with which humans can talk or interact, the field also encompasses knowledge representation, reasoning, and learning. Machine learning is then an important subfield of AI, and it involves supervised learning, unsupervised learning, reinforcement learning, etc. Over the last several years, many people have started to conflate AI and ML, but for someone knowledgable in these fields, the distinction is clear. ML is the practical application of algorithms towards taking inputs and producing an output prediction, and it's this area that contains the vast majority of jobs in "AI". Some basic applications of ML include spam filtering, face detection and recognition, product recommendations, fraud detection, revenue forecasting, gait and step detection, voice recognition, etc. If you look at that list and think about them, you'll come to realize that you've probably been consuming ML results for the last five years or much longer. If you want to work as a "ML engineer" in this area, you'll have to be knowledgable with ML algorithms, setting up data pipelines, running experimentation, and using ML software, such as scikit-learn, R, Caffe, Tensorflow, etc.

I manage a ML team at a large company. Let me make clear: Unless you have a strong academic background in this field, no one will take you seriously. I recently applied to be a principal engineer working on an AI/ML personal assistant, and the recruiter told me straightforwardly that the hiring managers are not interviewing anyone unless they have a recent PhD related to deep learning. I'm a bit elitist about this as well: I tend to turn away candidates that don't have at least an MS or PhD in a field related to ML. Why is this so? Because you need a rigorous background to understand why and how ML works, and this involves understanding loss functions, gradients, training vs. validation error, decision boundaries, optimization, and other things. You need to understand these things because a lot of current ML involves choosing the right knob settings (hyperparameters) that make your ML work best. If your ML algorithm isn't working well, how do you fix it? That's where this rigorous background comes in handy.

Now, there are many things related to ML that you can still work on if you don't have a strong background. As opposed to a ML specialist, there are plenty of positions related to data engineering (e.g. setting up and maintaining huge data pipelines), infrastructure administration (e.g. installing and mastering all aspects of Hadoop and Spark), visualization (e.g. creating dashboards that take fresh data and display it), among many others.

Re:AI / ML is not a hobby

[SlashTom]

Wow. That is pretty much the only serious answer so-far ;-)

I think you are right in saying that AI/ML is not something you pick up over a long weekend. Still, it should be possible for someone to give it a shot without doing a PhD.

There is a really nice Coursera course I can recommend: https://www.coursera.org/learn... If you follow that and make all the assignments, this will take you some months (assuming you don't do it fulltime), but you end up knowing a lot more about ML then a lot of others. Also there are tons of books that cover Machine Learning basics. Then, if you want to get your hands dirty, you can go to Kaggle and participate. I think if you do that for a while, and prove that you are really good at it, a company/recruiter looking for an ML person might be interested even if you don't have a degree in an ML-related field.

And as to the original question, what kind of job titles? I noticed that many times "Data Science" openings have something to do with some kind of machine learning. But the term is >>very As to "toying with Tensorflow", this is of no use whatsoever, IMHO, if you don't know anything about machine learning. So, first learn more about that, and then see how Tensorflow fits in, is what I would suggest.

Anyway, long story short: go for it!! You won't be an expert in a month, but you are in for a treat ;-)

look ma, no code!

[epine]

AI gets to the point were it solves a set of previously unsolvable problems, the algorithms are then researched and better non-AI solutions are then used to solve the same problems. Then AI falls out of fashion for a while and computer power increases thanks to Moore's law. Then it all repeats.

This old story is such a crock.

I highly recommend the following to anyone who wants a different perspective on modern ML:

* Talking Machines: Remembering David MacKay with Philipp Hennig — 21 April 2016

* Probabilistic-Numerics.org

This is plain old numerical methods, optimization, and search viewed through a Bayesian inference filter. I would never have termed any of this "artificial intelligence".

It took the recent large advances in unsupervised learning, the kitty classifier (and progeny), and the LSTM machine translation models to finally justify rethinking academic labels. Programs like SHRDLU from 1968 were perhaps explorations in AI, if our baby-step microscope is especially well focused. But this was closer to natural philosophy than what later became physics. Even our shiny new LSTM language models remain weirdly proximal to Searle's Chinese room. What have we really learned from watching our machines learn? Not a whole damn lot.

I'd nominate a term such as I-cubed: inexplicable inductive inference, or perhaps MIII: massively inexplicable inductive inference.

Even so impeded with an appropriate name, MIII is pretty mind-blowing. But it still ain't AI. It might be a viable building block to proceed in that direction, sooner rather than later, as we begin to erect dynamical systems upon this foundation. To drive the point home, it remains way overblown to call it MIIR: massive inexplicable inductive reasoning.

An Alberta AlphaGo Pioneer Is in China to Watch the AI Wallop Human Opponents

"Before AlphaGo, much of the fundamental games and machine learning research was done here," Muller wrote in an email. "If you look through the references list of the AlphaGo paper in the journal Nature, over 40% of these references have a University of Alberta (co-)author. Then, DeepMind greatly surpassed all of these previous efforts with their new ideas."

I haven't waded through this yet, but I suspect even the vaunted AlphaGo has a backbone of techniques that I personally wouldn't have classed as "AI" (or even AI-ish) by my own standards.

For decades, the big idea in AI was supposed to be recursion. Perhaps human language is recursive in theory, but it's only barely recursive in practice (nest more than three levels, your accurately attentive audience grows thin). Winograd is not completely wrong about this, but my long suspicion is that recursion is not going to enter the AI building through the ground floor.

Lately, we really have hit home runs with distributed representation, and to a lesser degree with convolutional image recognition. These are actual AI-ish ideas. However, two solid take-home techniques do not a field make.

Here's another possible intermediate term: generalized gradient exploitation (GGE). Plus there's tons of great mathematics about overfitting and regularization. But should we really call all this math "AI"?

In practice what AI ends up meaning is "look ma, no code!" Hey, we just built an impressive system without hiring rooms full of code monkeys, so we must be doing something right.

AI is not the moving target of lore. It's mainly our long AI pretension that fits the bill.

How many legs d

Re:"New field of AI", indeed

[religionofpeas]

That's like saying, "there's no story, just a book full of letters", or "there's no intelligence, just a skull full of chemicals", or "there's no weather forecast, just computers running code".

I started doing ML stuff as a hobby...

[raftpeople]

and tend to agree with your post based on my "non-expert but considerable time spent" experience.

My story:
I wanted to create intelligence via artificial life and evolution. I didn't want to create human intelligence, just tiny little creatures trying to survive type intelligence. I provided them basic sensor inputs and motor/movement raw materials but didn't program in any usage of those things, they need to figure out through evolution (how to see, move, find food, avoid getting eaten, etc.). They started with random neural nets and through generations increased in capabilities up to a plateau.

Some conclusions:
1 - There is a lot of foundational stuff I had to learn slowly and piece meal by googling etc., for example, what is a neural net doing mathematically, why/when would you use one.

2 - I thought I might figure out something interesting or clever - but the reality is that people with math backgrounds (e.g. PhD ) are the ones that are going to figure out that next clever insight. For example, someone once asked me why I was using neural nets and not support vector machines, so I read up on support vector machines and they seemed to be doing the same thing (function approximation). But I didn't have enough training to fully grasp why a support vector machine and neural are different, what are pros and cons. Reading papers online with teacher is a slow and cumbersome way to acquire knowledge in a complex area.

3 - Interesting issue not really related to the topic at hand but fun to talk about: My other conclusion was that guiding an evolving system towards intelligence is a very tricky task. My creatures hit a plateau of behavior that was at least interesting (chasing/tracking other creatures to get food, avoiding getting eaten, avoiding obstacles) , but difficult to get beyond. How would I need to change the conditions in the environment to push them beyond that into more advanced behaviors, for example hiding around a corner or hunting in packs, etc?. The initial environment needs to be favorable for guiding random brains towards some basic functionality, but then it needs to change and continue changing to keep pushing the evolution process towards more complex capabilities. That is a tricky problem, knowing which environmental conditions would reward intelligence over speed or strength or other attributes.