Machine Learning thread

Do NN require more linear algebra than knowing what matrix multiplication is?

I have this dataset where I have to simultaneously do:

>Correctly classify each observation into one of thirteen possible classes
>Estimate a certain numeric quantity Y for each observation

In other words I have to simultaneously do regression and classification. Each observation has 6 attributes, and so I need to figure out a regression model + classification model based on these six variables. From fiddling with the training data it looks like there's some pretty hairy interactions between the classes and the Y values. Another major problem is that the sample size isn't large: I only have 90 observations total, and some of the classes have as few as 5 examples.

So far, I've made a simple classification model that hits about 80% accuracy, but the regression model is really throwing me for a loop. Obviously I can just do regression on the entire dataset, but this gives pretty poor results. On the other hand I could do a separate regression model for each category, then use the predicted category to choose which regression model to apply to test data. The problem with this is that some classes have fewer observations than variables, which is obviously bad.

Anyone have any ideas on how to do this?

how can she slap?

I suppose you could get by and pass the class, but if you actually want to use ML in a non-classroom setting you need to know about vector spaces, eigen-whatevers, spectral theory, matrix decompositions, etc. If you already know matrix multiplication but you want something more rigorous, I recommend "Linear Algebra Done Right" by Sheldon Axler.

Wait I just re-read, if you're doing a masters related to deep NNs you definitely DEFINITELY need a solid understanding of algebra. Absolutely non-negotiable if you're doing it to that level.

>understanding of algebra
understanding of LINEAR algebra, jfc I'm tired

This machine learning bubble is gonna go down so fucking hard.
I mean, it's a cool concept.
But technology is just not there atm.
Only companies that can afford Tens of thousands of GPU/APICs has any chance of developing something useful.

See, machine learning looks like magic.
You just have your server farm run thousands of simulations, and it ends up behaving somewhat like you want it to.
Except, there's never a 100% chance it will.
It will just do it 'most of the time'.
You could train it more, but then it will never get to 100%.
In fact, there's no way to determine that a machine learning is 'safe' to operate in the real world.
What's worse is, you can't debug it, and see what went wrong.
Mark my words, people will die in autonomous cars, and we'll never know why.
Then they'll be pretty much banished, even though actual humans are way more dangerous behind the wheel.
It's just that insurance companies won't allow it.

There's no career in machine learning.
Unless you're a machine yourself.

>Will a career in machine learning/software development get me a comfy job

yes

>when the automation revolution takes the world by storm

don't hold your breath for that, the rhethoric squirted around the media landscape is overly optimistic about the actual state of technology because journalists are not engineers, they're journalists.
don't get me wrong, automation will progress in much the same way as it has since its beginning two hundred years ago - at a slow, yet accelerating rate - but it won't come to anything resembling a "revolution" over the course of our lifetimes (if ever).

you're right in your analysis that control systems operated directly by trained neural networks are not a reliable concept. your assumption that that is what is happening, however, is wrong.

neural nets are being used for the thing they've been designed for: analysing data and grading results (i.e. providing a confidence together with the result). the actual control systems are still somewhat conventional, human-designed pipelines relying on several sources of data analysis results, but on no single one alone - in other words, if confidences of results go to shit and/or stages of the pipeline fail completely, you will first see graceful degradation before full scale failure occurs. what this means is that, yes indeed, these systems will perform provably better than humans and will hence be allowed.

what you're completely right about though is the scalability of this technology, which in its current state is simply absent. google in particular is making huge efforts to combat this and they have made admirable progress, but the historical walls of computational power and data availability persist.

I'm not sure what you're talking about.
To me it seems like development stage error spotting.
Your autonomous car wont be 'learning while driving'.
It may send the data back for the next version, but it won't improvise in unknown situations.
For starter, it won't identify it as 'unknown'.
It will just carry on and do whatever the fuck it wants. Most likely won't end well.
That's a meager consolation to the family of the deceased that the data from their loss will be used to make it better.

Why not? Isn't that what machine learning is?

Well, again, not sure what you're talking about.
Are you talking about the car learning while driving?
No fucking way it carries the supercomputer needed for that.
Even then, it's trial and error, so it would just kill people, and learn that it's the bad behavior, in this situation.
But embedded systems don't do any of that.
They just have the neural network, and run it without checks.

But the software can receive updates.

Well yes, it can.
After people died.
Hence why I say autonomous cars will be banished.
Insurance companies won't be able to put blame on anyone.
This, after their income has been diminished to fucking nothing by constructor selling safe as hell cars you don't even need to drive.

Comfy job?
Not for long.

What machine leaning? All you need is an EE degree and knowledge how to program PLC.

Asking here since it doesn't warrant its own thread
Should I study Mathematics with AI or Computer Science with AI for undergrad if my main goal is PhD in ML?
Need to decide by next month

I studied computer science and mathematics and a lot of the mathematics side was really relevant to machine learning. I'd probably look at the specific modules offered

I don't think there is anything relevant to ML in CS beside data structure, algorithms and theortical CS which is math. If research in ML is your ultimate goal I would go with Mathematics.

math

>when
it's literally happening right now cunt

>/adv/ - Advice.
>implying those people go to college

this. the only board with higher level shitposting may be Veeky Forums