💾 Archived View for dioskouroi.xyz › thread › 25010114 captured on 2020-11-07 at 00:56:25. Gemini links have been rewritten to link to archived content
-=-=-=-=-=-=-
________________________________________________________________________________
I would not advise anyone to go down the "data science" career path at this point, unless one of these topics is a true passion of yours and you can't imagine doing anything else (and even then I would recommend alternatives if possible).
The explosion of AI/ML/Data Science teams in places that they really don't belong is going to have major backlash soon. The market is currently flooded with desperate PhDs who have reduced years studying their field into a flurry of buzzwords on a resume that they don't understand.
The vast majority of AI/ML/DS teams don't know what they're doing. Management and "leadership" in this industry would be laughable if it wasn't so horrific. The vision of AI/ML/DS has turned into a nightmare of people shoving data they don't understand into models they don't understand and making sure someone else is using that so they can justify their work. You could likely cut entire 100+ DS teams at most companies today and only see a positive impact on the business.
There are of course really interesting niches that require these skills and skilled practitioners. But if you're reading a guide on how to be an "AI Expert", they you likely aren't one of these people. And because of the signal noise problem in this space it's hard for even really excellent people to get matched to the hard problems that genuinely need their help.
There’s a lot of truth in this, but I find the people that pivot toward practical software engineering skills early in their PhD are still highly valuable to tech companies.
I think the issue here is you can’t become highly skilled in software engineering overnight, and compensating by flexing softer “problem solving skills” isn’t enough to compensate. If you want to be effective in data science, you need to commit to being a rockstar programmer AND statistician AND be well versed in a number of different algorithms.
Unsurprisingly, this is difficult, and so the market of people who do it well is small. That’s not, however, a referendum on the value of predictive modeling and software engineering. I don’t see advanced math and software engineering skills going out of fashion anytime soon, simply because they are hard to acquire and immensely powerful in the right hands.
I would say, however, there is ample opportunity for infrastructure designers/engineers in the space, especially in a world where everyone is going "cloud" but cloud doesn't cut it in the most demanding applications, so those of us who love data center design are sort of loving the resurgence of onprem/Colo usage and the opportunities for pushing the envelop on data transfer/processing rates, etc. The skills and knowledge gained are very useful in other industries, with a side bonus of, if you pay attention, knowing a bit more about the "big picture" of AI than many of the phd's!
I wonder what the impact of easy-to-use cloud ML tools is going to be on the industry. BigQuery already lets users train models using a simple SQL query. AutoML makes simple ML tasks as easy as run and verify accuracy and run again. Of course, a basic amount of knowledge is still need to avoid common pitfalls - unclean data, overfitting, etc. Perhaps this will just mean data scientists can produce better work with the same amount of time. It doesn't seem like a good use of everyone's time to have to know so much cutting-edge research to solve run-of-the-mill business problems. Hardly any other area of CS requires that.
For readers who read that and think they still want to do it:
> The vision of AI/ML/DS has turned into a nightmare of people shoving data they don't understand into models they don't understand...
Don't be this ^. Please.
If you do want to come into the field, please first take the time to understand the tools. Learn enough to re-derive and re-implement the tools you're using. Learn the assumptions that need to hold. Don't just learn to point the tools at data and hit go. It's not like a lot of programming where it either succeeds or fails. Failures are often silent in this field. If you can't reason through how and when a given tool will succeed and, probably more importantly, reason through how and when a given tool will fail, you haven't learned enough to know what to use and when.
I totally agree with you, but let's be honest, in this publish-or-perish world in academy, and the cut-throat environment of VC-funded start-ups, people will, if not outright lie, exaggerate the nature and importance of what they are doing. Too much money and "prestige" is at stake.
The "right" way is slow, lonely and unglamorous. And most importantly without that fast money, acquisitions and promotions.
> The explosion of AI/ML/Data Science teams in places that they really don't belong is going to have major backlash soon
I'm not disagreeing with you, just curious, did that happen with Big Data teams? If not, what happened with all the people working around that concept a few years ago? Might be an indicator of what could happen to DS teams.
I have heard hiring managers explicitly saying that data engineer role on your resume is considered a negative point for them.
Every time i fire a data scientist the earnings grow
Perhaps I may be mistaken, but this seems to be a very long road for a more shallow understanding of deep learning. I'd venture this was written by someone who has a more traditional machine learning background that wants new people to the industry to have that same foundation; however, I'd venture that that is a rather inefficient way to get to deep learning proficiency.
If I were to give a recommendation, it would be this -- pick a topic/project that interests you, follow the classic knowledge-bootstrapping process where you read through papers and (hopefully) have an expert or trained person walk you through the specifics, then get hands-on instantly. Especially with something like fast.ai that values empirical results over hard theory, something understandably popular in the field.
From there I'd recommend branching out, but I'd use a JIT approach. The field isn't necessarily super well-founded at the moment, and while machine learning fundamentals are useful, ultimately it's a waste of time compared to the long-tail benefit of getting immediate empirical results and feedback hands-on.
Just my 2c, YMMV, and anyone is totally welcome to disagree as they wish! :)
Best of luck,
T
> I'd venture this was written by someone who has a more traditional machine learning background that wants new people to the industry to have that same foundation
This is the vibe I got as well. Which is fair enough (to each his/her own), but I thought I'd mention fast.ai which takes the opposite approach:
> Harvard professor David Perkins, who wrote Making Learning Whole (Jossey-Bass), has much to say about teaching. The basic idea is to teach the whole game. That means that if you're teaching baseball, you first take people to a baseball game or get them to play it. You don't teach them how to wind twine to make a baseball from scratch, the physics of a parabola, or the coefficient of friction of a ball on a bat.
As for our opinion (which is just that, an opinion) why we think that the statistical foundations and knowledge about more traditional algorithms is important, it's based on the business needs and our experience. While it might seem less necessary if the goal is to "learn deep learning", it is highly relevant if your task is to "solve this business problem".
Our perspective is the industrial one. And while there are certainly many complex business problems where deep learning is required, there are more cases where a traditional approach is sufficient and actually the better solution (e.g. due to the memory footprint, latency or other reasons). We routinely work on both kinds of problems on a day to day basis, but we would never go straight to deep learning approaches if simpler and faster traditional methods comprise a better solution in a given use case. So, our employees are expected to know both and to be able to judge when to apply which approach.
Yes, this makes sense. I'd suggest retitling the article to be something along the lines of "ML Expert Roadmap", and then continue to flesh out all avenues. As it stands, the roadmap has really nothing to do with AI at all, but your point about not just jumping to the shiny hammer certainly rings true and makes sense. I certainly think that's the right approach algorithmically, especially if you're looking to be a more generalist data shop.
In a world of senseless marketing hype, I think it's a good idea to take the high road on this one. Reputation alone, even if less-buzzy words like ML are used in favor of AI, really carries a long ways. Plus, we're nearing the disenfranchisement hump, and AI's going to start taking a negative connotation with many businesses, I believe. Just shoot straight and I firmly believe it'll carry you for a long ways, there.
I agree with your suggestion, and that is what I did to get started in machine learning as well. Solving some kind of problem and doing practical projects and playing around with models is probably the most efficient way to learn. I started with the Machine Learning course on Coursera by Andrew Ng. And while going through, I started using what I learned on a small project. I did document my journey in case it's useful or interesting for others:
https://embracethered.com/blog/posts/2020/machine-learning-b...
"I'd venture this was written by someone who has a more traditional machine learning background that wants new people to the industry to have that same foundation"
Yeah I get the same feeling. I am hesitant to suggest it, but gives a little bit of a gate-keeping vibe. Like only once you've payed sufficient homage to the same education path that I took, shall you be granted permission to be a 'real data scientist'.
That said, I think there is value in something like this more for hiring managers than practitioners... which makes sense considering the intro comments.
You are right that the "Deep Learning" section is rather shallow up to now. We are currently working on expanding it to offer a more comprehensive view of the field and expect to release this update next week. Stay tuned! :)
With all due to respect to the author of the site, mastering all the materials in the machine learning or data scientist path will make you a solid 'applied' machine learning/statistical learning practitioner, but not an expert, and definitely not a research candidate.
We should be careful with how we guard our scale of semantic meaning - if somebody with an undergrad understanding of statistics (frequentist statistics only in this map) can be called an Expert in a statistical/mathematical field, what do we even call somebody with all the same applied software engineering & exploratory analysis, and business experience, but also a PhD in theoretical topics (math, stats, etc)? A 'super expert'? What do we call Francis Chollet, or LeCun, or anybody else? What's the differentiation between an expert from the roadmap and the team of individuals deploying GPT-3 into Google Assistant? Are they the same?
As a hiring manager and team lead for a large fintech firm in London, I would happily see an individual who had really mastered the above path(s) as a strong candidate for an intermediate or upper junior role in applied data science/machine learning. But ... it's not enough to be a senior, and certainly not an expert. Just my two cents.
Using a PhD as a gateway into applied ML is so horrifically misguided I hardly know where to even begin debunking it.
PhDs are _one_ especially crappy way to prove you have the intellectual chops to engage with ML. There are far more direct, practical, and expedient alternative paths to get there.
Importantly, the number of people who were perfectly capable of doing a PhD but chose not to (because, frankly, it's a very bad deal) vastly dwarfs the number of people who stuck around in academia and obtained one. Additionally, my observation at several major tech companies is that PhDs have a bent of mind that is roughly orthogonal to the pursuit of real business value.
Yikes! This doesn't really come across as a nice comment (bent minds???). There are good concrete reasons to pursue a PhD (ignoring soft reasons like pure interest): wanting a research career is one - it's pretty difficult to get hired as a scientist without a PhD. Also, historical evidence doesn't really support your claim that R&D is orthogonal to business value. Sure, pure science is often independent from $$ (despite plenty of examples of producing real value), but applied R&D is oftentimes parallel to value generation. if R&D in general is useless, why do top tech companies spend big money on research groups?
Getting a PhD is a fine deal if you have good reasons.
I explicitly caveated my statement with "Using a PhD as a gateway into applied ML" for a reason. The vast majority of people going into applied ML are not pioneering new methods. They're using ML as a tool to support business objectives. This is the group I'm talking about.
The phrase "bent of mind" roughly implies "the way someone thinks". Its usage is declining I suppose but there's nothing connotatively nefarious there.
Yeah - I looked it up. Thanks!
Bent of mind is a phrase that means proclivity or predisposition, and PhDs are famously arduous.
Expertise is relative. Having compiled a few kernels in my life I wouldn't consider myself an operating systems expert, but to a career plumber, I am. Heck, a friend of mine is a PhD in physics, operates a particle accelerator for work, and to him I'm an operating systems expert despite my protests to the contrary.
You're an expert if other experts say so. The opinions of clueless folk are just random noise.
Just because the man who can see is king among blind men doesn't mean they're qualified to call him visionary.
> What do we call Francis Chollet, or LeCun, or anybody else?
Ehhh, French? I like Chollet(many people dont) but he is not in the same league of Yan LeCun at all.
It seems both the author and you have a naive view of expertise in academia.
Agree 100% with you but would add that from the perspective of a lot of business people, the bar for expert is actually pretty low and however far over it you are, they dont actually get any more out or you because they don't know what to do with you. What a lot of businesses want is someone they can call an expert but that still operates within the realm and understanding of a non specialist manager.
At this point I feel like some crazy person talking to themselves when I say this, but "AI" is a research field that includes many more sub-fields than machine learning, let alone deep learning, and it's impossible for someone to be an "AI Expert" while ignoring this, even if they think they are _because_ they ignore it. I appreciate that synecdoche is a thing, but to recommend a "roadmap" to becoming an "AI Expert in 2020" while ignoring most of AI is...
... well I don't know what it is, anymore. Sign of the times? Perhaps the field (AI that is) should abandon its name just to be sure that if a backlash ever comes against deep learning it won't take everyone else's reseach with it?
On the other hand, I feel a bit like the horses and the cows in Animal Farm, when the pigs took over the Revolution. Not that there was any revolution, not really, but the way that the research trends have shifted lately, from what would make good science to what will make you hired by Google, is a little bit of a shock to me. And I stared my PhD just three years ago. It's come to the point where I don't want to associate my research with "machine learning" and I don't want to use the term in my thesis, for fear of the negative connotations (of sleazy practices and shoddy research) that I am concerned might be attached to it in the time to come.
And it's such a shame because the people who really advanced the field, people like Joshua Bengio, Jurgen Schmidhuber, Geoff Hinton, Yan Le Cunn etc are formidable scientists, dedicated to their preferred approaches and with the patience to nurture their ideas against all opposition. The field they helped progress so much deserved better.
Maybe a better way to attack this problem is to watch Harvard's Introduction to AI course (or similar one) and from there pick an area that piqued your interest. This is because AI is such a vast field that nobody can learn everything (on top of data processing etc.). I made some notes from the ^ course + listed relevant Python libraries to have a good place to start:
https://stribny.name/blog/2020/10/artificial-intelligence-in...
I think this would be better titled as machine learning expert roadmap. How come there is no mention of logic or knowledge representation?
Agreed. I think this is a well-intentioned effort at marketing and sharing knowledge but it seems a bit fluffy for the "AI" side of things. For ML, I think it's a pretty darn good foundation with the traditional hand-jammed/artisinally-crafted/exceedingly data-driven approaches.
Fascinating blind spot isn't it?
I really appreciate the taxonomy and links given there, but whenever I see someone claiming "you can't do Z before you've mastered X and Y" it tells me that someone's afraid of others succeeding with simply skipping the masterhood of X and Y and going straight to Z, picking up just enough X and Y in the process.
In fact, since the direct path is more rocky than going via X and Y, it naturally selects the most motivated people, so those succeeding in it will generally outperform those who went X-Y-Z.
This is great in some aspects but really shallow in others.
In fundamentals, I don't see a reference to geometry or calculus. Good luck understanding methods like UMap or more modern clustering techniques, optimal transport methods in AI, and more recent gradient techniques without those.
Some might say this is more for the AI researcher, but people need to implement bleeding edge stuff all the time.
I assumed a tech article and first read "AI extracts roadmap" like
or
https://github.com/Microsoft/Open-Maps
;)
Are people really planning their carrier like this? Isn't it more that you either work for some company nearby, or a company you know by some luck or explore the AI/ML topics you or your professor likes and you got stuck? (or you start disliking it and try hard to find something else)?
Nothing wrong with this roadmap but I'd suggest that this and similar ones are squarely in the "trade" category of formation, where the focus is on a large number of practical topics instead of a more solid grounding in the fundamentals, that mostly dont even concern themselves with practical dat science - what I would call the "university" approach, but maybe not in the sense of a modern university.
For work as a tradesperson, it is worth only focusing on the practical application- and I'm saying this genuinely. But I think it should be clearly distinguished from the different kinds of lasting benefits that a more solid fundamental education provides, including the flexibility to adapt.
I'm very biased here, having studies Electrical Eng and CS before modern ML was remotely mainstream, and comparing my conceptual understanding and what I learned in school about math and linear algebra with the way things are understood by tradespeople with far more knowledge of modern tools than me. So crotchety old person- maybe, but I'm happy I went to university.
The number of highly controversial opinions on this, to me, only shows that it makes sense to try to map a potential path. While the criticism in the comments here may be valid, it is mostly unconstructive and without any proper explanation or reasoning. It may not be perfect, but it's a great overview of what to look into.
There's a split happening right now in the market between Data Scientist and Machine Learning Engineer roles. The former is generally a more soft-skills heavy role with results that are meant to drive human decisions. The later is about building systems that automatically make decisions (ie: recommendation systems, etc.). Data Science is over saturated while ML Engineering is less so.
This should be renamed to "Intro to AI Roadmap"
What's on this roadmap will get you in the door with some of the basics. In no way will it make you an expert.
Awesome resource.
A quick plug and reminder that a learning roadmap for data engineering that is detailed with resources is available here:
https://awesomedataengineering.com
For those who wish to focus on data engineering.
I'm off the main hubs, SF, London, NYC, but here, in my gamma level city people talk a bit about AI but I have yet to see applications that are not vaporware. Don't get me wrong, I love ML challenges, but the real world seems to have so many other little things to solve first, before we can finally realize our ai dreams.
Things like Alexa or Google Translate already very much exist, use modern AI techniques, and are not vaporware. That's why so many corporate research labs - basically all the FAAMG companies - do AI work.
Surprised to not see fast.ai or some of the more popular courses on here as well to gain more experience
If you are looking to work in an enterprise as a data engineer, submission of research papers with code is certainly not a pre-requisite. Learning something useful like CPMAI methodology is more practical.
Why is "Big Data Engineer" not a child node of "Data Engineer"?
Yeah that seems like a major flaw in this taxonomy. Like does the author really think there is significant enough difference between 'data' and 'big data' that it warrants an entirely separate track? That's silly.
I chuckle a bit when they put PCA as being the very last thing you learn, way after dimensionality reduction (PCA is the classic and most "simple" dimensionality reduction technique)
Also, the reality is that very few full on AI experts (and yes, this means Ph.Ds with a lot of publications at top conferences who are also FAANG applied ML engineers) will have more than 70% of these skills. Luckily, this is a field which highly rewards specialization. They don't need the breadth if they have depth in their particular areas.
My experience has been that in the vast majority of applied roles involving AI, there is not enough emphasis on the software engineering skills necessary for productizing the models, leading to situations like baron_harkonnens critique.
I do appreciate they mandate paperswithcode.com