I just completed ST101's uberhard programming assignments :P

Though I must say its useful to write functions for those base statistics, because you can use them while answering the in-video quizzes

@IvoFlipse i started that course but then i just didn`t like it ... is it good? is it worth catching up?

Its probably not worth it

Or at least not yet, especially if your math knowledge is up to date

I like that he explained Maximum Likelihood Estimators to me, because they never bothered in PGM

oh then i need to watch it ... I had to read a lot to understand MLE but still i am not sure i get it entirely

Well what he did was calculate the probabilities of throwing a dice

Then you pick the side that has the highest (most likely) probability

If only they explained everything in such simply terms :P

oh really

well I liked the explanations but the initial videos were way too basic

anyways you are saying it involves programming too ?

That was the 'last' assignment

which used 2 out of 6 exercises :P

actually this is good

So no, its not challenging at all

Physics on the other hand actually does have some nice questions, though I guess it depends on how well your trig knowledge is

to program

No, sadly not

I wish they added some programming assignments, because that would save a lot of students a ton of work

too easy ?

is this different course?

Yeah, PH101 and ST101

oh ok

btw Coursera has a very nice course

social networking

starting in sept

its all networks

I'm watching the Udacity version of that CS215

my friend told me it will be good and that I need to revise my linear algebra for that

wt`s in CS215?

algos?

Yeah, focused on networks

In the video I'm watching now he's explaining a connected components algorithm

if IKEA made particle accelerators alexislloyd.tumblr.com/post/1611284936/…

oh i need to watch that

i am slowly entering into system`s biology and networks are core components

I'm watching it again, because I wasn't paying attention enough :P

thats nice

Hmmm I should try and create a graph based on my paw data

Then I can use connected components to track paws in consecutive frames

will this be separate for every single dog or similar ones?

@GauravJain Let me rephrase it, if the paw makes contact with the plate, it activates pixels. Then for every frame that particular paw is in contact, it keeps activating pixels, mostly in the same area. If you visualize this by stacking up the frames you get a 3D structure where the paws location is defined by the pixels it activated. Those are all connected within the same frame and with frames in the consecutive frame. So I could run a 3D version of connected components on it

oh thats pretty cool

@GauravJain Well I haven't figured out what data structure I'd need for it

I believe it already exists, because they use it for MRIs, but I couldn't really grok the algorithm :P

how about a simple linked list

But linked in what direction?

Though I think I haven't tried thinking the problem through enough :P

a 2D connected component is just a contour in OpenCV, so I'd assume if a contour is a direct neighbor of another contour in the next frame, they're connected

then it would be a linked list of contours I reckon

yup

Or rather of coordinates within the contour, like a lookup table

frame wise

as soon as you get a new activated frame

then you can merge the similar frames to shorten the list

The good thing about a search like that is that for example with a human foot, first we see the heel, then the mid/forefoot come down, then the heel takes off and eventually the forefoot too

The heel and forefoot are connected during midstance, but aren't at the start and end. Using a connected components search, I would find that they are connected at some point in time and therefore are connected

yeah thats true

Hola the computer vision class just posted a bunch more video's

@IvoFlipse lol I already forgot about that class

Need to start watching again

Well now there are actually some video's

and the latest batch is about object recognition

nice

He posted a comment that he underestimated the work that was required

As may of you have noticed, over the last month my production of lecture material has been rather slow, for which I apologize. The teaching of this course is a strictly unpaid volunteer effort on my part, and that of the TA, Bharath Hariharan. At the same time, I have many other responsibilities and duties such as my research and obligations to my employer. Part of the problem is that my guess of the time needed to prepare online lecture material was a gross underestimate.

While I have been teaching in a classroom at UC Berkeley for many years, the online medium required the creation of almost all the lecture material from scratch, which took quite a bit of time. I was forced to choose between quantity and quality and I chose to sacrifice quantity.

I was forced to choose between quantity and quality and I chose to sacrifice quantity. My thought has been that even if it is at a slow pace, it is still serving a useful purpose, and the content produced so far will provide a good base for subsequent versions of this course. Some of my Coursera colleagues whose courses are much more polished than mine have had the benefit of these rounds of refinement, whereas the current course is my raw, first attempt. Sorry!

wtf. Is he trying to make us feel bad?

No idea

there's more text after it, but not really easy to paste in here :P

@IvoFlipse Yeah, I'll read it later

@Phonon @IvoFlipse Ha! I had completely forgotten about that class

Rightly so @Mohammad

I didnt like the way it was going anyway, reallly reallly slow

@Mohammad exactly my point = )

Now I know why hehe

@Mohammad I was always listening to it on 1.5x speed

@Phonon @IvoFlipse At least he had the cojones to come clean tho

He could have said that much earlier though

@IvoFlipse Yeah, before we forgot about that class lol

I almost think they shouldn't start a course until they've prerecorded the material

@IvoFlipse I thought thats how they did it -

Unless they have the time and resources to produce it quickly enough like Udacity does

@IvoFlipse Most of the stuff like homeworks and such is usually handled by class TA who volunteer. Maybe this guy hadn't had any? Don't know.

@Mohammad well the downside of that is that they often resort to picking already recorded material (like OCW lectures), rather than making new stuff

@IvoFlipse Yeah hes pretty much part timer on that as he said. Whereas Udacity is full time rock-your-world education

OCW?

@Phonon You'd hope they already had plenty of this material from Berkeley's Computer Vision class, its given every year

Open Course Ware

Ah

But not every year by the same professor

@IvoFlipse I see

@IvoFlipse Yeah, new genuine content is really needed badly. Im still waiting on SOMEONE for an email. ;-)

@Phonon Its too bad, because its about the only topic I really actually need to follow right now :P

@IvoFlipse Same here more or less.

What classes are yall in ATM?

@Mohammad Algorithms, CV (lol), Udacity's CS101, also watching a bunch of mathematics lectures. Number Theory at the moment.

@Mohammad Waiting for Part II of Algorithms. I really enjoy that class.

should be offered in August

@Oh

@Mohammad All new Udacity classes and CV

And I finally started to try and apply some of the stuff the past 2 weeks

@Phonon @IvoFlipse Nice - I actually started to teach myself ICA - almost completed it.

ICA?

Independent Component Analysis

@IvoFlipse independent component analysis

Based on maximizing kurtosis of a distribution

very neat idea

@IvoFlipse en.wikipedia.org/wiki/Independent_component_analysis

@IvoFlipse It is VERY powerful I am seeing already

@Phonon Or minimizing ;-)

@Mohammad Really? I thought it was the other way around.

@Mohammad Oh well, have to brush up on that

@Phonon You can have both it turns out, if you maximize it, you can extract super-gaussian signals, (voice, songs, etc), if you minimize it, you can go more after 'simple' ones like a saw tooth signal, etc

@Mohammad From I remember, the more you mix distributions (mathematically, convolution) that more Gaussian you final distribution becomes. Kurtosis can be used as a measure of non-Gaussianity, therefore the more kurtosis you have, the less mixing.

@Mohammad interesting

@Mohammad Any particular paper? The original one from Helsinki?

I have no idea what that means :P

@Phonon Yes thats true. I think since it came out of voice field they focused on that, but it also turns out that say, if we have signals that are not sub gaussian, (negative kurtosis), (and know they are there), we can use that as a metric instead

@Mohammad Makes sense.

@Phonon oh dude, get this book - you will love it. Same for you @IvoFlipse (amazon.com/Independent-Component-Analysis-Tutorial-Introduction/…)

But it sounds like a way to decompose a signal into multiple parts of which its actually comprised

@IvoFlipse Exactly - except we extract them now based on their PDFs, not traditional time/frequency properties

@Mohammad Nice. Not expensive either.

@Phonon Yes, I HIGHLY recommend it. It is what I am learning from.

@Mohammad But where is that based upon?

@IvoFlipse Whatcha mean?

Those pdfs are based on the traditional stuff

Hmmm there's something in here that looks like Deep Learning

@IvoFlipse You got it. :-)

Damn, too bad my Google Books ripping site is offline

Its available there too

@IvoFlipse The awesome insight of ICA is that this - say you have a signal - any signal, and you histogram it. You will get some PDF distribution. Say now you add some combination, (linear combo) of a bunch of signals. And you histogram that.

@IvoFlipse Now as it turns out, the PDF/histogram of the bunch of linearliy combined signals, will tend towards a gaussian - look like a gaussian PDF, whereas each of the individual signals that made up the combination, is either going to be super-gaussian, or sub-gaussian. The measure of this 'how gaussian are you' is called Kustosis.

@IvoFlipse So ICA is the process by which you 'do something to data matrix' -> extract a signal, -> measure its kurtosis -> increase/decrease what you were doing to the matrix to extract signal, etc etc

Now this is exactly why I'm doing Stats

@IvoFlipse udacity?

because I have no idea what I would do with a histogram of my data

@IvoFlipse PCA (Principal Component Analysis) looks for emperical eigenvectors that have the highest variance, 'energy'. Decomposes data into orthogonal basis functions, (eigenvectors) that best describe the energy. ICA on the other hand, will decompose it into independent basis functions

Yeah

@IvoFlipse It started?!

@IvoFlipse I signed up, but didn't do anything yet. Do they have deadlines for stuff?

Week 3 already

@IvoFlipse F*ck lol

Nope, just an exam

so you could wait until the last week :P

@IvoFlipse Nice. Just like grad school! = )

lol

Now I need a coffee. brb

@Phonon Only because you don't have access to all the material in week 1

@IvoFlipse Got it.

@Mohammad I have no clue what the difference is :P

@IvoFlipse I'll start catching up with you soon then

Testing can be a pain in the ass though

@IvoFlipse Is the final on a specific date?

@Phonon The last week is the final I believe

or a long weekend

@IvoFlipse K.

I got screwed last time because my internet connection was killed

Now that we've moved, I've got a pretty solid/fast connection :D

@IvoFlipse Wow, that sucks

@IvoFlipse Nice = )

And my standing desk, finally!

Though I was completely knackered yesterday from moving

@IvoFlipse Standing desk?

@IvoFlipse Yeah whats a standing desk?...

@Mohammad He must be taking a picture right now ;)

I was looking for a video

@IvoFlipse wow

I only installed it yesterday, its not yet in a state I'd like to eternalize it in :P

@IvoFlipse Cool idea

@Phonon lool

I wanted a desk like he had, with a curve, but it wouldn't fit in the car

@IvoFlipse @Phonon Neat idea... now we can be like those people on startrek who stand up at their posts doing stuff on their computers! lol

And this desk surface turned out to have an ugly scratch, but it was so cheap I can't be arsed to go back to Ikea :P

@Mohammad That's the idea, to not site down all day long

Interesting

@IvoFlipse @Phonon Have you guys ever heard of a 'polyphase FFT'?

@Mohammad I've heard of polyphase filters.

@Mohammad Not sure if it's from the same area

@Phonon Ya me too...

@Mohammad Probably not

@Phonon @IvoFlipse Hmm, check this out about 'polyphase FFT' dsprelated.com/showmessage/45449/1.php

blog.dropbox.com/index.php/new-dropbox-pro-plans

@Mohammad Interesting

@Phonon Seems to be yet-another-way of creating an STFT

@Mohammad Too bad there's a link in there that doesn't work

@Phonon Yeah.

@Phonon That entire site is really quite old, and from the main characters on it, I bet theyre all before our time.