/sci/ - Science & Math

[Anonymous No.12362118]

How much do you actually need to know to actually get a Machine/Deep Learning job?

I have pretty extensive experience (as a hobbyist) using most of the Python AI meme libraries, but my degree is in engineering. All the job postings I see always ask for CS, math, or engineering degrees.

So what is the actual dividing line here?

[Anonymous No.12362141]

[Anonymous No.12362144]

>>12362118
Just train a neural network to forge degrees.

[Anonymous No.12362160]

>>12362141
Oh trust me, I'm totally aware of this.

The only reason I'm asking is because these companies are listing all these non-CS degrees and it makes me wonder if it's viable.

[Anonymous No.12362165]

>So what is the actual dividing line here?
Think about the sort of degree the person processing your job application will have.
>Oh, my future boss will read my application!
Business major. You're in.
>Oh, some random HR person will read my application!
>Hmm, it says here Anon uses computers... to do machine... computer.... uhh... math... I think that department does math on computers too?
You're in.
>Oh, my future colleagues will read my application!
>Looks like he has little experience, but most of the required background. Let's give him a couple of interviews.
Half way there.

[Anonymous No.12362166]

I want to ________________________________
Margaret Hamilton.

[Anonymous No.12362169]

>>12362141
>visualizes data aesthetically in Jupyter Notebook
>Copies code directly from stack exchange and github
Apparently I'm a chad.

[Anonymous No.12362235]

>>12362166
What would you do to her?

[Anonymous No.12363781]

It depends on the jobs you're applying to/interested in. I don't know what a "Machine/Deep Learning job" is. If you want a job in research (in a decent lab) you'll need a PhD and top conference pubs. If you are just looking at jobs that apply machine learning models the qualification requirements go way down. Most non-research machine learning jobs are just going to be building data pipelines and data engineering. These type of jobs don't require extensive machine learning knowledge.

[Anonymous No.12364116]

>>12362118
I mean, it depends on what you qualify as an ML or deep learning job. Do you mean writing queries to popular libraries? Do you mean designing and implementing learning architecture, which is a full time engineering job? Or do you mean doing research into nontrivial use and developing it more, which is a mix of theory and engineering research? These routes all involve different levels of academic aptitude.

[Anonymous No.12365350]

>>12363781
Non OP here, say I want to go the hard route and become a researcher, specifically what kind of degrees would I need?

[Anonymous No.12365360]

>>12362235
Considering she's like 85 I think she aged fairly well

[Anonymous No.12365672]

>>12365350
I am a researcher in ML for mechanical/materials engineering applications (I have already published 15+ papers in the JCR).

To be honest, in my experience you need a Ph. D. in a specific area to understand fairly well how to apply ML and complex numerical procedures to either solve problems or craft real hard knowledge. Also, aside from degrees, it is usually more important to have fully grasped some disciplines, just to mention few: Statistics, Probability, Tensor calculus, OOP programming, Numerical Methods (including hard optimization methods), scientific writing (to effectively report advancements and scientific reports).

It is a long path, but if you are really into, it pays every second of the time you invest in it.

[Anonymous No.12365704]

>>12362118
ask in HR maybe

[Anonymous No.12365734]

>>12362141
I'm Chad but using Arch Linux.

>>12362160
The thing with pure data "scientists" is that they have no idea about what they're applying their algos to. Most companies need someone fluent in ML but also real-world problem solving. As an engineer, you know about safety factors, stability of structures, reliability, and generally what's important for making something work in real life. That's what they're looking for in those cases. Someone with a good foundation of let's just say physics, plus knowledge of how to apply ML to solve those problems.

[Anonymous No.12365756]

>>12362118
Everything you've learned in your degree is based on theories and logic. As soon as you spend some time in a job, you will realize how worthless your degree really is.

[Anonymous No.12366605]

I do deep learning research focused on pharma applications. My personal background is a MS in chemical engineering. From my experience it's easier to get into the industry if you work on applications you have a specific background for (ie when I was new and less experienced I could still run circles around CS grads who don't understand SAR). If you want to do research at a major lab and don't want to spend 5+ years on a PhD, it's easier to get a data science or machine learning engineering job and then move into research.

For doing stuff on the job, the most important thing is being good at practical applications. You should be able to read a paper and implement it from scratch with good quality code. There's also a lot to be said for knowing how to architect good ML systems (the whole infra/serving system, not just the model algorithm which on its own is rather useless) and choose the right ML projects based on business strategy.

For theoretical background, you need to know linear algebra, calculus and statistics as applied to ML. The key here is you don't need to know everything about all those fields. Some people get sidetracked trying to master everything in linear algebra. A good reference is the first section of the Goodfellow book (www.deeplearningbook.org/) which is basically a review of linear algebra, calculus and statistics applied to ML. You want to have enough foundational knowledge that you understand why you're doing certain things. For example when you use Kaiming initialization to initialize your weights, understanding why that's a good thing to do.

Overall though the most important thing is being able to put ideas to code. If you're serious about pursuing this, you should spend more time coding up examples than studying theory. Don't start trying to code everything from scratch, you just get lost in the weeds. Start with high level libraries and peel back the layers until you get to a from-scratch implementation.

[Anonymous No.12366748]

>>12365672
I know that people usually do not go into research for money or fame but does it at least pay reasonably for all the extreme effort you have to put into it?

[Anonymous No.12366887]

>>12366605
Alright, I'll check out the math section. I've already taken general courses on all the math you touched on, but nothing for the sole purpose of ML. And I guess the 'from scratch' part is where I am lacking the most. I understand the general concepts, but I am still very much dependent on python libraries like tf, keras, and pandas.

[Anonymous No.12366908]

>>12362118
How the fuck am I supposed to know about ML as a math major?

[Anonymous No.12366916]

>>12366748
Not that guy but ML salaries (at least in the Bay Area) are on the same level as SWEs.

[Anonymous No.12367337]

>>12362118
>i have eng degree
>jobs ask for eng degrees
where is the problem?

[Anonymous No.12367342]

>>12366916
i thought a lot of ML people are just SWEs now
even then at worst a lot of places will hire you as an swe if you can't find or don't want a job in research

[Anonymous No.12367368]

>>12367342
>i thought a lot of ML people are just SWEs now

I wouldn't agree with this statement. Things are still evolving pretty quickly, but ML Engineering and ML Ops are becoming their own thing.

There's a lot that goes into building a production level machine learning system beyond the algorithm, and setting it up right required ML knowledge that SWEs generally don't have.

A lot of it is related to automating the ML algorithm lifecycle. You need systems to actually serve your model, so that involves building auto-scaling inference systems and monitoring to catch regressions. You also want to be constantly collecting new data, processing it, retraining models (either online, nearline or offline) and serving new models. This entire system should be automated. And of course you never have just one model, you have this same process and automation occurring over dozens of different ML algorithms.

There's of course overlap with traditional SWE and devops, but ML deployment has its own unique set of problems, tooling/infra and skill set.

[Anonymous No.12367373]

>>12367368
To add an example, here's a simplified schematic of the Netflix recommendation system. It takes a lot of infra to actually turn ML models into a product.

[Anonymous No.12367396]

>>12367368
>ML Engineering and ML Ops are becoming their own thing.
that's kind of where i was getting at. when staff are expected to write their own code or do their own operations, or end up in one of those positions because they have ML on their resume, they might end up working on the support rather than the core problems

i work at a place with extremely high ops turnover. i'm a developer by title but i work near ops so much that i feel like i'm killing my career staying here

[Anonymous No.12367541]

>>12367337
>Job Requirements: EE or CE Degree
>go to Intel interview
>no VLSI experience

b-b-but the job posting said I just needed the degree

[Anonymous No.12367563]

>>12362118
phD from a top school with IPS/FAIR/etc publications
t. grad student with lots of friends in the job market

[Anonymous No.12367582]

>>12367396
I guess it depends on where you work. I've always been at places that had a distinct break between ML research and ML engineering. Also I'm used to "ML Engineer" being its own title separate from SWE tracks. However these are all larger companies with more mature ML products.

[Anonymous No.12367596]

>>12362118
>watch sage maker autopilot tutorial video
>trial and error factors until model is more right than wrong
>deploy sage maker model
>copy paste aws Lamda code from stackoverlfow for data pipeline, model integration, etc
Look ma I'm a full stack data scientist

[Anonymous No.12367611]

>>12367596
have fun getting a job at FAANG research doing just that
that only happens at shitty startups

[Anonymous No.12367648]

>>12367611
>FAANG research
you do realize the people doing AI/machine learning research at FAAANG are literally the best of the best right? The kind of people who have their pick of positions in academia, hedge funds, etc.

[Anonymous No.12367669]

>>12367648
It's not that hard if you apply yourself. Furthermore, why compete if you aren't competing to be the best? I'm in a seminar with FAANG scientists right now and it's not that hard if you apply yourself

[Anonymous No.12367724]

>>12367669
Sure as long as your realistic about it, it's probably the equivalent of a professional sports person in terms of chances of employment compared to say a regular code monkey at FAANG. A lot of the FAANG researchers are the kind of people who drift between academia, tech, finance, startups and anything else that takes their fancy, i.e. they are good enough that they have options.

[Anonymous No.12367734]

>>12367611
FAANG research groups represent a tiny sliver of actual ML/DS jobs

[Anonymous No.12367776]

>>12367734
>>12367724
Imo if you're not aiming to achieve excellence and be the best you can possibly be in any field there's no point. I'm talking about achieving your own potential, not societal definitions of success. I'd get bored out of my mind if I were at a job where I was only using 50% of my capacity

[Anonymous No.12367826]

>>12367582
my personal experience with devops is that i really, really want to get as far the fuck away from working with deployment automation, integration tests, oncall rotations, or log diving as humanly possible
this is the only SWE job i've ever had though so i don't know if software development itself is miserable, operations is a shit field, or if i'm just a miserable person myself

>>12367776
>Imo if you're not aiming to achieve excellence and be the best you can possibly be in any field there's no point.
in academia maybe. a lot of people just want to live a comfortable industry job because they really don't think they could make it trying to compete

[Anonymous No.12367841]

>>12367582
>I've always been at places that had a distinct break between ML research and ML engineering.
i was implying what i feel like is happening is that people who want to do research are being moved to other roles just because they need to fill them

[Anonymous No.12367849]

>>12362118
>How much do you actually need to know to actually get a Machine/Deep Learning job?
From what I've seen, the ability to lie convincingly to people that the crap code you wrote is actually 'intelligent', regardless of it's output being total nonsense. Basically: a degree in Marketing.

[Anonymous No.12367871]

>>12367776
Sure but I think your missing my point about being realistic about job/position competitiveness. Reality is a lot of people who go down this path end up as well paid code monkeys cranking out reports and CRUD apps.

[Anonymous No.12367882]

>>12367826
>this is the only SWE job i've ever had though so i don't know if software development itself is miserable, operations is a shit field, or if i'm just a miserable person myself

I think it's the field. Anyone I know who has done DevOps or Data Engineering has hated it and moved to a different field. I think it's because 90% of your job becomes "AWS Plumber" which fucking blows.

[Anonymous No.12368018]

>>12367882
i mean my job is better than most ops work in terms of workload

it feels arrogant for me to say but i feel like i'm not fighting real problems, but instead all the hard stuff is because there's a bug in the system. everyone who left my team said something similar

[Anonymous No.12368745]

>>12362118
Try tensor calculus without library.

And also Java.

[Anonymous No.12369017]

>>12365672
What degree would I need to work in a company like openAI?

[Anonymous No.12369074]

>>12367871
It's also a misalignment with what people think is cool (cutting edge research, etc) vs what is actually done in industry (use boring methods to solve practical problems). 90+% of models used in industry are either linear/logistic regression or a decision tree variant. Most companies applying ML don't have research teams because they don't really benefit from it. You don't need to hire a ton of MIT/Stanford PhDs to make a spam filter or figure out what factors are driving sales growth. Even the engineering aspect might not apply because tons of companies that use ML as a data analysis tool don't need production serving infra.

[Anonymous No.12369460]

>>12365672
>I am a researcher in ML for mechanical/materials engineering applications
Could you please please please tell a bit more about this? As a mech. engineering graduate currently pursuing a Master's in CS, it's extremely relevant for me.

[Anonymous No.12369463]

>>12362118
You need to know how to turn a computer off and on again, the automated scripts and batch jobs implemented by the previous engineer has already been amalgamated into the A.I. provisioning service and is at your disposal.

[Anonymous No.12369466]

>>12362144
Why? What humans are we trying to trick and to what end does tricking them aid anyone other than bury people in paperwork ad finitum?

[Anonymous No.12370203]

>>12362141
>use cython
>achieve nirvana

[Anonymous No.12370227]

>>12362118
I didn't even know you could get employed for deep learning shit, I just thought it was basically a cult of people with a collective open secret of wanting to make AI waifus real.

[Anonymous No.12370231]

>>12365672
>tensor calculus
this isn’t a thing, and you retards don’t know math.

[Anonymous No.12370243]

https://www.youtube.com/watch?v=9SAZ-9OVt2M&t=3s

[Anonymous No.12370260]

>>12370227
>I just thought it was basically a cult of people with a collective open secret of wanting to make AI waifus real.
honestly i really feel like it's full of people who don't intrinsically care about AI but just want the prestige and money of pioneering a new field