/sci/ - Science & Math

>>10387239
>hating machine learning because a bunch of ignorant mathtards think they can do it
Mathfag detected.
CS majors just ignore them. They'll never accomplish anything so nothing wrong watching them waste their time

https://blog.piekniewski.info/2018/05/28/ai-winter-is-well-on-its-way/
https://www.technologyreview.com/s/612768/we-analyzed-16625-papers-to-figure-out-where-ai-is-headed-next/

At least share the relevant links. Machine Learning is dying out because it has been stagnating for the last 2 years. AI will focus on a completely different approach in the future. And we are entering a AI winter as we speak.

>>10387239
so you can tell me, that u are e.g. able to do image recognition without a CNN, and you are abel to find the convolution kernel for the objects without the help of nn?

>>10387279
Maybe for research, but in industry there is a lot of room for growth. Many companies are still only beginning to use ML in operations and there is going to be a lot of opportunities for data engineers.

l>>10387239
ML math is extremely trivial, simple lin algebra. Fuck me cause I wanted to escape the no-math hell that is CS and study my master in ML, but seems thats brainlet math aswell so ill have to get a CS master instead :(
>>10387649
yes industry will need ML people, just cause research isnt making groundbreaking improvements every month doesnt mean ML is dead for those who work in industry.

>>10387649
>>10388102
I work in the ML industry for about 3 years now. People are getting cut and most companies are scrapping their ML projects and wings.

It's the other way round. Governments and universitites will be okay. Businesses like google, Nvidia, Tesla, Facebook will be the ones that will fire their ML departments or downsize because there will be no real developments anymore.

>>10387270
ym except invent math that crackpot ML people poorly graft onto their models as they realize their methods, which they have no understanding of, hit the wall the rest of us knew they would?

>>10388102
I have a background in theoretical linguistics and math, and so many people have asked me why I didn't just do NLP/ML, and this is basically it. When the NLP/ML people are more confused than linguists, who are already half retarded, i'm out. There is more to math than misapplied linear algebra.

>>10388270
not to mention the state of it now is a dead fucking end and there's no convincing these people because they are the kind of people who think that a vector is an array of numbers.

If we're going into an AI winter, does this mean that pretty much any other science are in their winter too ?

>>10388141
>Businesses like google, Nvidia, Tesla, Facebook will be the ones that will fire their ML departments or downsize because there will be no real developments anymore.
The big tech companies are only a small piece of the pie. There are many areas which will benefit from ML which are only beginning to implement it.
For example, a CTO at a local mining company told me they started using ML to optimize road gradients to improve efficiency and to predict when machines will fail and what parts to order in advance.
There are lots of opportunities in the economy for these types of improvements using existing methods.

>>10388141
Can you give some examples of what projects you've been working on?

>>10388916
Some very dumb fresh graduate stuff that is boring. And working with a pharmaceutical company to use ML to determine the likelyhood that a new drug will pass FDA regulation.

I'm now looking to getting to work on either a government project or at a university because the industry is going to be hit hard over the next year.

>people think that X is the big new thing, therefore X is not a thing at all
machine learning is useful and interesting. it's been around for a long time and applications are now becoming feasible. the fact that it's easy/hard/math-heavy/math-light isn't even worth talking about.

>>10389235
Please actually look into machine learning some more especially the progress it has made since late 2016 (which is none) to understand why it's getting in a dangerous position where expectations of investors aren't met by the actual performance.

>>10388234
yes, handcrafted image classification algorithms are totally coming back

Machine learning is a method for producing function approximators. That's it.

>>10389305
>Machine learning is a method for producing function approximators. That's it.
Agreed, but that is pretty useful.

>>10388270
>>10388275
>waaaaah I wanted to be super special
virgin mathfags with their custom models btfo by CHAD programmers with linear algebra and a rack of GPUs

>>10389310
Of course it is, I am not denying that. But it seems to me like many brainlets ITT had some very misguided ideas of what ML is all about.

>>10389312
probably just gamers mad that their graphics cards are more expensive now

>>10389318
>gamers
more like GAYmers

>he doesn't "grow" optimal networks using genetic algorithms
not gonna make it

>>10387239
There's still a 6 math problem that is not solve yet. Why don't you solve it using excel.

>>10389146
Whats your education background and how much math do you use on a daily/weekly basis? Do you just learn the math behind ML during education and then completely disregard it when you get a job? Aka do you write equations and shit if you work in ML or is that just when you get your education?

>>10389146
These are bad news for everyone working in The Cloud, DevOps and Big Data too considering we're collecting data to be used for ML later on right?

>>10389498
Data collection is going to continue or even increase because they are valuable and will be sold to other firms even without AI.

>>10389403
Depends on what you do. Some people actually change the ML itself by fucking with the nodes but most people don't do this anymore. Instead you tweak the way you feed information in the neural net and decide how much nodes you are going to use etc. It's just fancy data science after all. The math involved is actually pretty trivial but that is not where the difficulty lies. The difficulty lies in adjusting the neural-net to be as optimized for the situation as possible. Different data needs different feeding methods etc. What works for playing chess doesn't work for object recognition.

>Things that will stay with ML
Optimizing ways to feed data on a case-by-case basis. Optimizing the neural-net itself for a specific task
>Things that are disappearing
Tweaking the mathematics within the neural-net itself. Backpropogation. Administration and monitoring of the neural-net. hand picked (curated) training of the neural-net.

Basically 80% of what the job is is going out of the window as well as most research in the field switching to other AI approaches means most people are going to lose their jobs. AI winter is all but guaranteed.

>>10389520
What for? What's the point of collecting massive amounts of data other than analyzing it?

>>10389520
After reading this response it's quite clear you have a very tenuous grasp on ML. It's doubtful you even work in the field, and if you do you're a low level intern at a company who isn't ML focused.

It's sad that people here will read your responses and not pursue an interest in the field because there's a wealth of opportunities for people with even basic analysis skills nevermind the ability to apply models to data and program. Yes deep learning is over hyped from the perspective of general intelligence, but it also has the ability to drastically change hundreds of fields, mostly clerical in nature(so it's likely to automate jobs that require a degree like paralegals rather than manual labor). If all we accomplish in the next five years is merely an improvement of current technology then it will still have great potential and economic viability.

>>10389569
I oversimplified it because I don't want to give a lecture on the subject and make it easily graspable for people that might want to go into the field.

>>10389569
It's even sadder for people to get into a field because universities shill for it on a constant basis, racking up debt to get a degree on said subject (because let's be real, it's hard to get a job on ML without a degree, not impossible but hard), and then get canned because the hype is dead.

>>10389569
How much of the ai winter is a sci meme and how can i be sure? theres a master programme in ML at my school and im heavily considering it because i dislike the 0 math at the CS master programme. But i would want to use math when i actually work, not just simple import tensorflow as tf.
ive googled people with the same degree and many have the title "data scientist" and "machine learning engineer" but i have no idea wtf their worktasks are. Someone mentioned on their linkedin that they worked as a "ML engineer" at spotify and were working on the multi-armed bandit problem.. sounds interesting but how do i know if thats the norm or just a one off?

>enroll in ML class
>teacher warns us about the fact math is primordial for ML and the approach will be rigorous
>homework is just pajeeting around in Python
>textbook explains nothing and is overall incomprehensible
>a good quarter of the midterm is MCQ whose questions were copied form the previous years' midterms
This will shape the future, boys.

>>10390480
You know what you must do.

>>10390480
>loser at no name university takes course for brainlets
>complains that ML isn't rigorous.

Can you solve these without looking at the answer faggot? This is the most basic ML that is learned.
https://github.com/zyxue/stanford-cs229/blob/master/Problem-set-1/ps1.pdf

>>10388482
this, it's filtering down to small-medium sized companies now

>>10387279
Why does nobody attempt to do machine learning in pure C, Im thinking of doing so for fun and starting from scratch instead of relying on libraries

>>10390480
Which textbook was recommended to you?
>>10390574
Also this, I had the same experience

>>10390942
Good luck datamining.
The only thing C is good for is parsiny bytes, actual data it's horrible. It's doable but why would you torture your mind with managing memory.

Just use R or mathematica like any Statistician and Data-miner.


t. computer scientist

>>10391094
Also that ML-winter is on its way is bullshit. It has just started, perhaps it is stagnating on the researching side but when it comes to practical uses it is still at the very beginning.

in 1990-2010 everyone made websites and that was the big thing. 2010-2016 everyone made apps and everyone wanted apps. Nowadayds it's the clouds, clusters and machinelearning, everyone wants to apply neural nets and shit.

>>10389325
>He burns a totally unnecessary amount of compute instead of using transfer learning to get the same performance in a fraction of the time

>>10390462
>ive googled people with the same degree and many have the title "data scientist" and "machine learning engineer" but i have no idea wtf their worktasks are

It's hard because it varies hugely from company to company. They could be anything from data bitch to high level research.

>>10387279
>70's
>researchers create machine learning
>see how useless it is and drop it
>fast forward 40 years in the future
>dumb millennials reinvent the wheel
>meme it because it can play monkey videogames better than them
>no real application besides quirky programs
And the cycle repeats itself.

>>10391100
The way I see it, we're past peak hype but we're not even close to peak application/development.

We're finding out that a lot of the lofty promises/speculation over deep learning for AGI and other bullshit are going to fall short. At the same time, application/use of data science, traditional ML and deep learning is only increasing. More and more companies are realizing they are sitting on a fat pile of data that they can turn into money.

Going forward I expect ML/DL to become both less sexy and more useful. It's not Google Brain/self driving shit, it's grocery stores, agribusiness, shopping malls, banks, and other shit like that. ML/DL packages will become more abstracted to the point where everyday employees will use it the same way they use excel.

>>10391190
Thank for your words. That was what I was trying to say.

Test

>>10388270
>I have a background in theoretical linguistics and math

>>10389325
>maybe if I say 'genetic' people will think i am smart

>>10391431
THE AI IS TAKING OVERRRR

>>10390942
because that sounds cool until it takes you 4 years to scrape data you could've gotten in 7 minutes by typing install.packages into your R console

I'm not a robot

>>10388102
>escape the no-math hell that is CS
>master in ML
?????????
Why didn't you do either math or theoretical CS in grad school? Those fields are *explicitly* all hard math. For the record, theoretical CS isn't doing sipser level problems in automata, but things like lower bounds using lots analysis and difficult complexity theory problems.

>>10387239
>>all math
All convex optimization. Not really all math. There's some calculus in there too, but that's about it. ML is more or less rebranded statistics.
>involves no real programming
I like programming, but why is this a downside? All the libraries or either there or waiting to be written

ML is a meme because it's not a very convincing of its correctness or even methods. It's a meme because it's experimentation in CS without theory gone way too far, and if we continue to trust its results, we also put blind faith in its failures that we cannot yet understand / perceive. Also everyone and their mother wants it for no work

>>10391451
yes u r

There was a kid in my undergrad capstone project who wanted to pick a project and implement machine learning. Myself and the rest of the team quickly said "fuck that" and picked a different project. He left our group to join the group that picked that project. They did not succeed.

>>10390942
Writing linear algebra functions that actually work, run fast, and can be put on a GPU is really fucking hard. The rock bottom mechanics of most ML/DL libraries were written by Korean geniuses who literally invested a lifetime of education into optimizing that exact task. Don't try to reinvent the wheel. Just say thank you and import the library.

>>10391633
>experimentation in CS without theory
nigga there are literally decades of research and theory on this shit.

>>10391748
>We have been shitting for decades and have become extremely proficient at it

>>10391629
because i already have a BSc in cs and dont want to completely switch fields to math. and theoretical CS sounds boring. I dont want to work in academia.

>>10391796
>I'm too much of a brainlet to read the theory so I'll assume it doesn't exist

>>10391443
>not randomizing the weights of your massive rnn until you get a strong AI
or a fuckload of NaN

>>10392691
>avoid the math in CS
>avoid math
Why are you complaining about CS being mathless when you avoid the math in general? What’s your idea of a math-y job such that you don’t complain about it being “mathless?”

>>10393218
>or a fuckload of NaN
>When your RNN is so good the probability of incorrect words is negative infinity

>>10391748
>>10392908
Again, most of the theory in ML is inherited from optimization, some calculus, a bit of analysis, and then the rest inherits very directly from statistical methods. I've read ML theory. It is literally the efforts of stat departments to stay relevant after having their field mature in a way that doesn't promote new research all the time.

When I say that ML (and AI in general) is lacking in theory, it's that we have little in the way of generalizing our results to talk about more than just than one domain at a time. In that regard, we "barely" know what we're doing, and the trend of current ML research is to just to apply it to more domains. Barely anyone is interested in making more sense of it, and the theory papers that are out there on the subject tend to draw weird glances that paint the stereotype that ML researchers don't know math.

It's not inherently a bad field, but in many regards, it is easily one of the least exciting and most volatile applications of basic CS in its current form.

>>10387239
>all math

>>10387239
Most people in machine learning have a jobs which are based on implementation and programming. The mathematics is done by PhDs

>>10393468
Holy shit.

>>10393821
Breaking News: CS students are brainlets

>>10390942
Google darknet and blas

>>10389520
What should people do their PhDs in for big bucks if not AI? Also, from what I can tell there does seem to be some growth in using AI to assist creating in new materials (see: citrine.io)

The warped perspectives on this board are always hilarious since few if any actually are employed
>get my Machine Learning diploma
>get to work as a Machine Learner
>clock in on my punch card when the steam whistle blows
>start a 10 hour day of Machine Learning
>my department starts losing money
>get laid off, can't do anything else because all I know how to do is Machine Learn

I don't care what company you work for, this isn't how real life works. I work for one of the large tech outfits, we use it effectively in many cases, but the people employed to be involved with use of it are experts in the system they are trying to improve, not machine learning itself. The hard part is figuring out which questions to ask to best utilize it, what are you trying to do, what information do you need, how can you get it, how can you most efficiently make use of it, how to structure the logic in the analysis of the data, etc.. VERY few people are actively involved in the math aspect of it for 40 hours a week+. Specific degrees mean fuckall anyway once you've entered the workforce.

>>10393831
I think this is unrelated to CS. I've been seeing much more of this for the past year than ever before.
I feel it's a symptom of 10000 third worlders getting an interrnet connection every hour.

>>10393831

College students in general are like that. I've heard people with a master's in engineering and coursework in statistics say "well, 6 has been rolled a lot this game so there's little chance it will be rolled again" too many times.

1. be a chad programmer
2. create more bugs and complex code
3. job is secured
4. helps future programmers to get hired to maintain the ugly legacy code or create a new one
5. future programmers repeat step 2

1. be a virgin statistician
2. do glorified statistics that is ML
3. import tensorflow as tf
4. job has a shaky foundation
5. employer does not need any of his kind
6. regret & distress suddenly appear

>>10391162
Based

>>10387284
You can do image recognition work knn buddy

>yfw you find out that linear regression and spreadsheet solver is "machine learning"

>>10393384
i dont avoid the math, i just dont want to drop cs and do 100% math.
mathy job = job that requires you apply university maths, like calc, lin algebra, graph theory, time series analysis etc

>>10391162
>>10391443
>let me "guess" the network architecture
this is where the complexity arises and why the winter cycle happens

>>10390574
>UC
>no name
Try again.

>>10391090
pic

>>10387239
>tfw physics/biochem double major with 100th percentile MCAT applying for a ML job at a neuro lab
>competition is stanford CS major who started his own app company and is already 3rd year at #3 med school
>i have zero experience with coding or ML
>PI and i have an interview where we talk about literature, etymology, steppe people, and hippos
>sees that my previous research experience was in particle physics and cosmology
>points out my MCAT score being higher than anyone else he's seen
>hires me ASAP tells the other guy he'll give him some project

i spent 16 hours a day for a month cramming ML and learning python for this job. i actually fucking learned this shit. my PI is a godsend for believing in me.

mfw

>>10395908
Good on you anon

you can tell machine learning is a meme by the amount of women that "work" in it

>>10393958
most people here are butthurt undergrad brainlets in the 130-150 IQ range who are mad that the niche they occupy in society is rapidly dying. Increasingly every problem can be solved by non-brainlets (150+) with the aid of automation or simply sheer numbers - now there are hundreds of thousands if not millions of them - or copious amounts of subhumans (<130) working with tools developed by non brainlets. Nobody needs mid IQ stamp collectors anymore.

>>10394126
I am a third worlder Indian, and all these things are taught to everybody vigorously till the 10th std.
After which we have to choose a branch, in which we want to do high school diploma. Like in science, commerce or arts.

>>10394266
They are just being superstitious, they don't mean it ;)