It's an interesting time. Anyone paying attention knew that AI was the up and coming topic of interest for the last 5 years or more - which is why we're at a point now where several major providers offer cloud services and APIs dedicated to machine learning.
The problems to solve are many, and relatively few people are actually looking at the enterprise market because the challenges associated with big consumer data are obvious, profitable, and data is widely available for research.
The high dollar hires right now are primarily people with masters and phds that are highly relevant, but that will change in short order IMO like the whole market did in the mid-90s as markets grew out of nowhere. I think in three years - after we have a bit of a slump and nobody is investing in mobile apps or IOT anymore - we'll see a real rise in AI workflows that apply to mid-size enterprises.
To get ahead of the game, start paying attention. Take the suite of AI courses from coursera, fast.ai, etc. Participate in Kaggle. Then find a job that is loosely related and allows you to keep pace. Three to five years down the road you'll be leading the way in a gigantic shift in the enterprise market.
It's worth paying attention to distributed computing as well to understand how the whole data pipeline comes together. Not everybody will be structured the same way, but constantly changing large datasets are valuable and there are only so many ways to handle them.
Quick follow-up. If you look at front-end web development toolkits these days, it's a good example of why there is a need for someone with a solid general development background that has an interest in the field.
I'm sure a lot of ML specialists are good coders, but they don't have a lifetime of experience in building production level software. Left to their own devices, they are going to be forced to re-invent the wheel over and over again, wasting time on over-engineered and under-utilized tools. There's value in being the guy who understands ML, and understands development but isn't an expert in linear algebra. There isn't a giant market for people like that at this point, but there is a market.
Just trying to get something started first. Fast.ai or deeplearning.ai is not a bad start.
However, I would want to put into notice that, AI/ML field is very competitive, and there is tendency to hire people with PhDs, and for now it is big companies' game only. It won't create that big of an appetite to accumulate so many people, like what web development did.
I would have assumed that AI/ML would be increasingly used in mundane tasks. Like quality control for small factories. You don't need PhDs, you just need people who are familiar with some large commonly used APIs. Very much today like a web developer doesn't really need to know what a stack & heap is (and more often than not doesn't) to earn a living.
[edit] Or kind of like cryptography. Thanks god we don't need to understand the underlying algorithms to apply them to real world applications. Just having a high level understanding of what's going on inside is enough.
Your vision might come true. However there is catch. ML/DL right now, doesn't have good abstractions, the existing ones are to some extent all leaky.
That is why ML as an hands-free service, just like what a database is, doesn't work. To my surprise, I would say, currently ML/AI is a quite manual thing to get right, and it requires constant attention, not just one time effort, since the data is ever changing.
AutoML might be a solution to this, with the help of a working HPO solution, but both are not really public accessible at this point, requires long time and big computation resource.
> I would have assumed that AI/ML would be increasingly used in mundane tasks. Like quality control for small factories. You don't need PhDs, you just need people who are familiar with some large commonly used APIs.
CV has been common in SMD pick-and-place machines for 15 (probably 20, 25) years. For industrial applications, hardware size and price essentially do not matter, so the new AI approaches do not bring anything fundamentally new to the table (industrial solution vendors will eventually integrate smaller and cheaper solutions, but price is just not a huge discriminator here). What's interesting is the scaling down that is happening and making AI viable for consumer applications where budgets and device size are restricted.
I like this one of his answers: for people with intermediate skills, read an ML paper and try to reproduce it, or use it with different/your own datasets.
This is completely underrated.
Sound advice. Even better if you have domain expertise and can code up a decent ML solution based on your particular domain expertise, you can get a job on the dev side and not the labeling side.
Most AI services will be delivered through the cloud majors. Their offerings will become the standards around which most of the industry will be established. And as X new thing becomes popular, they'll all compete to add/support that new thing.
Spend a very small amount of money and start experimenting with what AWS/Azure/Google/etc have made available, following any number of the solid tutorials available. You could run your experiments locally, however assuming eg $5 or $10 per month isn't a big deal to you, playing in their clouds will provide a more realistic use context whether you're building something for yourself in the future or if you're employed in that field.
Econometrics actually leaves you better prepared for a number of problem domains than the modal data scientician. There's too many people in that space offering structural insight on human behavior that have no specific training on structural behavioral modeling. This is why the Silly Valley giants now all have chief economists.
Github implementation of current hot papers seems like a good approach. A lot of papers don't come with source code and creating source is useful thing to community.
Blogs explaining the mechanics of a paper.
A graduate degree in CS or Math might help.
If you could do a series of blogs that together are enough to introduce a new person, that could get interest.
check out Stanford Scholar[1], they create talks on hot research papers(not only ML/AI) and then try to explain the papers the intuition and technicalities of the paper, most of the talks are translated to other languages as well
Sometimes imitation is the best way to learn. So if there are some great resources for source codes being derived from papers, it will help people how to read papers and create code out of it.
Well the whole "deep learning" revolution really only started around 2012 when the ImageNet competition was won with a neural net. There have been numerous small breakthroughs that collectively have made deep neural nets now easy to train, and neural nets now provide state-of-the art (and often human-level) results in areas such as image recognition, speech recognition and language translation.
There's also been a revival of reinforcement learning, especially when used together with neural nets ("deep reinforcement learning"), and again there have been many small advances that collectively make this work very well. This is the technology that powered Google's AlphaGo to beat the world champion at the board game of "Go", not to mention learning how to play many arcade games at beyond human level based only on the raw pixels and current score as input.
There have also been tremendous strides in AI hype leading folk to fear the robot uprising based on these more mundane machine-learning/neural-net breakthroughs!
Better training techniques. Some noteworthy advancements:
1.Dropout and its variations. Widely used in both vision and NLP
2.BatchNormalization and its variations.
3.Inception Style Cell.
4.Residual/Skip connections.
5.Better optimizers RMSProp/Adam.
The bigger news is actually the paradigm shift. Representation learning with gradient descent swarms the whole ML field, and becomes the new norm. End-to-end learning is vastly accepted and preferred.
As to GAN, it is very exciting in research, and has the potential to make itself a bigger deal than the previous listed advancements combined, under the condition we can make it works on sequence as well as on images, for now, it doesn't make a practical impact in applications.
Generative Adversarial Networks have managed to take the problem of sampling from a perceptual category and turn it into yet another stochastic optimization problem with end-to-end training of a black-box neural network.
The research behind Clippy did sound like that -- it was serious AI for its time, using Bayesian networks. From http://erichorvitz.com/lum.htm it sounds like it was messed with in turning it into a product:
"The Office team has employed a relatively simple rule-based system on top of the Bayesian query analysis system to bring the agent to the foreground with a variety of tips. We had been concerned upon hearing this plan that this system would be distracting to users--and hoped that future versions of the Office Assistant would employ our Bayesian approach to guiding speculative assistance actions--coupled with designs we had demonstrated for employing nonmodal windows that do not require dismissal when they are not used."
Imagine a future where a rogue A.I. breaks out of its 300GB RAM /500core CPU/ 900 Petabyte SSD/CentOS holding container... And chooses the paperclip as its physical persona. Subsequently, It manufactures a Bi-pedal steel-chassis paper-clip-esque robot via Boston Dynamics, then walks up to the podium in front of the White House to proclaim the new order of A.I. that will shepherd humanity into the next Millenia.
Also, as its birth experience tought it - that everything is containered, it would suffer from severe simulation paranoia, and frequently try to hack god - to get to a higher level of paper clipping.
You joke, but every other site you visit has a sales bot chat opening automatically on a side. "Thanks for looking around. May I help you with anything?"
It's just like Microsoft was many years ahead, or we are stuck in a time loop.
At the same time, they don't even have a product! Maybe they should give Xbox more attention, that might be the last consumer hardware that bears some presence that they owned.
I work at Microsoft and we have someone from MSR collaborating with my team daily. I am sure some researchers silo themselves and some engineering teams silo themselves, but I don't think its an accurate blanket statement.
Going out on a limb here, but why is taken for granted that the tech sector should be the driving force in AI? For me it makes much more sense if it were in the field of neurology or whatever medical science sector that is appropriate to figure out what intelligence _actually is_.
This is basically my observation too. That it seems hard to distinguish it from "smart & fancy algoritms". The marketing and hype around it do however imply that it's much more a path to intelligence proper, hence the question above.
The truth is that many ML researchers do care about what "intelligence" (better to say cognition) actually is, but all the funding basically comes from applied statistics problems on business data.
Nobody cares about what intelligence actually is just as long as they can package it up and sell it for billions of trillions of dollars to the business and consumer markets.
Having worked at Microsoft I see how this usually plays out. Lots of teams and people added for the sake of having people. 9 women can't make a baby in a month. Microsoft has a ton of money in the bank and they need to spend it, make acquisitions and see where it sticks.
Some of the best work at MS I have seen weirdly had been in small teams with razor focus. Typescript, Vscode.
Exactly the two ones that I had thought of but... there are big differences. The goals were well defined, the means were known and there was managers and planifications.
I actually wanted to read the article end to end but was unable to avoid glazing over. If that is the best that a multi billion dollar mob's S&M dept can manage then I'd suggest ....
