From ML class, what kind of math is this?

Want to study ML. Saw this in a ML lecture notes. What can of math is this?

Other urls found in this thread:

arxiv.org/pdf/1701.07875v1.pdf
people.eecs.berkeley.edu/~jonlong/long_shelhamer_fcn.pdf
twitter.com/NSFWRedditVideo

It's mainly probability theory.

t. phd student

interuniversal teichmuller theory

Also just wanted to point out that machine learning always looks like this. If you want to just learn some shallow theory behind how to implement a learning algorithm, do not take a machine learning class. Maybe an undergrad one isn't as bad but at the graduate level it's pretty hardcore.

Makes sense. I don't know any probability theory. Where should I start & what pre-reqs should I have before I start with probability.

There's a mix of probability theory, multivariate calculus and linear algebra (at least up to SVD). You need all of that before you can really start learning ML.

Computer science.

Isn't it
[eqn]
\begin{align}
\frac{1}{N}\sum\limits_{i=1}^{N}\log\frac{f(\mathbf{x}_i)}{p(\mathbf{x}_i)} &= \frac{1}{N}\sum\limits_{i=1}^{N}\log\frac{ \frac{1}{N}\sum\limits_{j=1}^{N}\mathbf{1}(\mathbf{x}_i = \mathbf{x}_j )}{p(\mathbf{x}_i)} \\
&= \frac{1}{N}\sum\limits_{i=1}^{N}\log\frac{\frac{1}{N}}{p(\mathbf{x}_i)} \\
&= \frac{1}{N}\sum\limits_{i=1}^{N}\left(\log\frac{1}{N} - \log p(\mathbf{x}_i)\right) \\
&= \frac{1}{N}\sum\limits_{i=1}^{N}\log\frac{1}{N} - \frac{1}{N}\sum\limits_{i=1}^{N}\log p(\mathbf{x}_i) \\
&= \log\frac{1}{N} - \frac{1}{N}\sum\limits_{i=1}^{N}\log p(\mathbf{x}_i)
\end{align}
[/eqn]

I don't see any computer science in that image.

You can never really tell with random ML excerpts because people introduce their own notation in every other paper.

harr harr, then why is it offered as a graduate CS course?

Well, that's pretty much all I did in my computer science courses.

It's applying mathematics for a niche purpose - computer science.

Let me give you an analogy. Would you call physics math just because physics uses math as a tool?

Then why would you call computer science math just because we use math as a tool?

Non-sequitur.

Right idea, wrong application. Statisticians use computers and computer programming as tools. That does not mean that statistics is computer science.

Statistics isn't quite ML. Which you would know if you ever took a ML class. ML uses probability theory as one of the main tools however.

Ok thanks.

This is a grad level CS course. Haven't seen any applications yet.

>Statistics isn't quite ML
Do elucidate the differences for us.

"""Machine learning""" is a buzzphrase moniker for prob/stats used to secure research funding, get dumb undergrads to take your class, and spam in your startup's bio to ensure that you're bought-out.

If that is what ML really looks like, then that looks boring. Who is really interested in that shit? There are more satisfying areas of math than statis/probability shit.

In statistics you study distributions to understand the structure and properties of them.

ML is about learning an unknown function given a sample from an unknown distribution. As you can imagine there is some overlap which is evident in the way we model the problems.

Some statistical techniques work well in ML, such as linear regression. Others are purely ML such as neural networks.

But what ML entails is exactly what I described, that you try to learn an unknown function of observed data.

This is the theory behind ML. If you want to gain wizard status then you need to understand the theory child.

Of course you can always use a library and try experimenting on real datasets.

yeah i guess ill take the effort to learn the theory. id rather do that than just use libraries

Are you just going to slink away now that your low effort trolling has exposed your ignorance?

Filthy undergrad.

Using libraries is a good thing and a lucky part of being programmers as well. It's nice to be able to actually see results and get some tangible motivation to dive deeper into the theorems.

>In statistics you study distributions to understand the structure and properties of them.
>I'm the ignorant one

quite true my man. i completely agree. you're getting me motivated again to learn ML.

Pray tell undergrad-kun, what do you do in "real" non-interpretive statistics?

Are you too pajeet to understand the validity of my generalization? It's okay, I understand.

Quantify uncertainty.

Very discrete
Much wow

You're thinking more of probability than statistics, which in either way is not the same as ML which I pointed out the difference for you.

It's funny how much bullshit people throw at ML who know nothing about it except whispers made by nervous outdated professors worried about their own funding.

Also fuck you. You shit on my perfectly reasonable definition of the problem of statistics and then you respond with the ***most*** undergrad-ass definition possible.

Statistics and probability are two sides of the same coin, "undergrad-kun".

Take a look at some current research
arxiv.org/pdf/1701.07875v1.pdf

Clearly by your own admission this had to do with statistics since it's concerned with the properties of a measure of distance between distributions. To prove these properties, however, the authors must delve deep into measure theory and Borel sets, topics typically categorized under fundamental probability theory.

>In statistics you study distributions to understand the structure and properties of them.

Ask yourself this child, does anything you say refute anything I said?

Yes "child" it refutes your insistence that statistics and probability are somehow different. And since your retard definition doesn't mention probability, you clearly don't understand either.

Oh and just to play the "lets google a random paper" game here's a paper on CNNs:

people.eecs.berkeley.edu/~jonlong/long_shelhamer_fcn.pdf

Surely this is just statistics right? Not worthy of it's own field of course.

How are distributions defined? As PDFs. Do you really need someone to spell out everything for you?

You asked me to tell you how ML is different than statistics, which I did.

After being BTFO you are now ass devastated and going off on bullshit tangents that have nothing to do with the topic at hand.

Another thing child, if statistics and probability weren't different, why are they two different fields?

>this will be entertaining

>How are distributions defined? As PDFs
rofl

As I thought you're just a neural networks fanboy that thinks shitting out yet another network architecture to chew on yet another labeled dataset to increase a benchmark by some small margin constitutes a major contribution and makes you a big shot.

Yeah, that's it's own field, it's called "garbage".

mmm the salty tears of an undergrad. Delicious.

Hopefully one day you will understand that being open minded to other fields is a sign of intelligence. I understand that undergrad is an isolating experience that's very hard and scary but learning to be receptive will be a great asset in life.