The h Bar

AccidentalFourierTransform

02:27

> im not emily please dont tell my mum

Anonymous

03:56

@AccidentalFourierTransform Lol

Anonymous

Link?

Anonymous

BTW long time no see ;)

David Z

05:04

@Blue It's deleted now

Anonymous

@DavidZ Ah, nevermind :P

Avantgarde

Hey

No one

Slereah

08:53

Another PhD application sent

undefined

@Slereah good luck :)

Slereah

10:14

Hm

There's a lab that does quantum gravity but it's LQG 🤔

John Rennie

I think choosing to work in quantum gravity of any form would be a brave career choice right now.

People have been beating their heads against it for decades with essentially nothing to show for their pain.

Slereah

It's okay, I don't mind not being useful :p

Slereah

10:44

Hm

Greek physicist in a french lab

Should I write the email in english

Secret

10:58

When in doubt cold calling european researchers, use English or Spanish

In other news, this week's NewScientist have a couple of interesting thoughts that is quite counter to my intuition:

1. About the LIGO gravitational wave business that should have been reported many years earlier, it appears the NewScientist team have been involved in uncovering irregularities in some data collected by research institutes. LIGO's response reminds of what the general public think of the government, but I trust they are actually doing things about it

John Rennie

The NS article is a storm in a teacup. Some researchers have attempted to re-analyse the LIGO data, failed, then claimed LIGO didn't observe anything. Which is funny seeing as how the neutron star merger was also observed optically.

But a result is that LIGO will publish details of how to do the analysis, which is good.

Secret

Indeed, that's how science keeps itself in check. Looks like we can benefit from these independent investigators in keeping our integrity in check. Most data faults are often accidental though, caused by human errors and other unconscious things

after all, it is really hard work to crawl through all that data, even with advanced AI

2. It turns out mass air travel has a mixed effect to pandemic spreading. Common wisdom said that air travel should result in pandemics to spread much quicker before a warning signal is evident, but the latest research suggested that because pandemics often begins with local strains getting some mutation and become deadly, and people infected with the less deadly ones can get some shielding effects from the genetically similar ones,
air traffic actually would have spread these mutants all over the globe, thus in a way vaccinate people globally making them harder to suddenly get all sick unl…

3. In regards to taming extreme weather,warming climate models showed changes the jet stream in such a way that it can triple heatwaves and floods. Interestingly, while cutting sulfur pollution from coal burning can stop this rise and stop arctic shrinkage, the decrease in the cooling effect by them on the mid lattitudes can make stronger heatwaves and storms there

4. And finally, the main topic of this weeks NS is that too much self control can lead to regret and other things

All 4 of these topics caught my attnetion because they all share this "It looks like it is X, but actually, it is Y and -Y" property

Abstractly speaking, being exposed to these problem solving scenarios and case studies can be useful in brainstroming solutions, because of the inherent complexity and that they are opposite (not just different) from our common sense. Most of our modern issues including the messy political sphere all have some traits of these and thus solution that can tailor to these "problem profiles" can be very effective, I think

It seems to me that the current trend of our society is that we increasingly need to employ solutions that go against our common sense, plus some suitable mixture of common sense to spice it, in order to untangle the problems and reach a solution

Captain Bohemian

12:11

I find when I read an email the first time, I may not take the message as exactly as what the mailer wrote.

If I have the chance to read it again and again, I may find some message I didn't take the first time I read it.

I don't know if others reading my email experience the same situation.

Secret

12:43

I sometimes overlook some phrases and then when reading again, become surprised that there are new phrases there

Slereah

Di Francesco is at Saclay

Aaaah

i'm so sending him a mail

Captain Bohemian

13:19

@Secret actually when I read message in this room, I sometimes experience that. So communicating on web is really different from face-to-face conversation.

so I think rebuking others on web may not be that effective as doing that in reality.

people may just overlook some of your sarcastic phrases.

also, I don't know how to express tones in wriiten English.

rob

@CaptainBohemian ♪♫♬?

4

Captain Bohemian

13:35

@rob I mean the tone of talking. When I speak, I can express my emotion in corresponding tones, but when I write, I can't. The same sentence with different tones often conveys different meaning, like good-willed or ill-willed.

rob

@CaptainBohemian I know. I was trying to be silly, but it didn't come across well because written tone is hard even for native speakers.

You're not alone.

user1936752

Sorry, I'm using the chat for the first time so I'm not sure if this is the right place but does anyone have thoughts on Sandu Popescu's recent note on the arxiv? arxiv.org/abs/1811.05472

Secret

13:53

@user1936752 no equations?

rob

> In this note I will present a subtle interplay

Too subtle for me, I'm afraid.

user1936752

14:09

Makes me question if I even understand the basic formalism of pure states and mixed states...

vzn

14:55

@user1936752 (welcome) hes a very prolific and diverse researcher/ writer. like some his attn to bell nonlocality, bohmian mech, interpretations of QM foundations, & collaborations with aharanov etc. yes there is something very subtle about the density formulation of QM. it naturally leads to realist interpretations. few have the virtuosity/ tenacity to pursue it to the end. arxiv.org/search/…

vzn

15:46

@Secret interesting! some links newscientist.com/article/… newscientist.com/article/… newscientist.com/article/…

LIGO is revolutionary new science and uses very new techniques/ methods of analysis. the detectors are incredibly sensitive and can apparently pick up earthquakes and maybe even nearby trucks/ traffic. the data analysis is a triumph of data science. did feel their process is somewhat opaque so far. LHC sometimes has similar troubles with false positives and very subtle statistical analysis. science in progress...

Space Otter

16:43

@JohnRennie What is the "norm" of the 4 momenta?

John Rennie

@SpaceOtter how much have you done on four-vectors in special relativity?

Space Otter

Nothing at all

THis was a 2d problem

John Rennie

@SpaceOtter ah, OK :-)

Space Otter

Probably should have said

John Rennie

In that case I'm not whether I should be talking about four-vectors as it might just be confusing, though if you go on to take any serious SR course you'll learn about them since they're pretty fundamental.

Space Otter

16:47

We just have momenta of two daughter photons in $i$ and $j$

John Rennie

And they make calculations a lot easier in many ways.

Space Otter

Okay sounds useful

John Rennie

@SpaceOtter Regular momentum is a vector $(p_x, p_y, p_z)$. Yes?

Space Otter

Mhmm

enumaris

whoop dam

Space Otter

16:49

Yes

enumaris

got asked a simple technical problem for an interview and I totally blanked

failsauce

John Rennie

@SpaceOtter and the magnitude of this vector is $\sqrt{p_x^2 + p_y^2 + p_z^2}$. Yes?

Space Otter

Yes

John Rennie

In special realtivity we are dealing with four dimensional spacetime not just three dimensional space, so vectors have a time component as well. So the four-momentum is $(p_t, p_x, p_y, p_z)$, where the time component is $p_t = E/c$ ($E$ is the total energy).

Space Otter

Where is $pt=E/c $ from?

John Rennie

16:54

In SR the magnitude of a four vector is called the norm. It's calculated in a slightly different way to the magnitude of a three-vector. For a four-vector the norm is given by:

$$ |p| = \sqrt{-c^2 p_t ^2 + p_x^2 + p_y^2 + p_z^2} $$

Note there's a minus sign in that equation.

Space Otter

I have a lecture I'll be back later, thanks though

Okay, but where is pt=E/c from?

John Rennie

@SpaceOtter That's complicated ...

enumaris

17:12

hmmm

feelin dumb today XD

Semiclassical

@SpaceOtter The following skips a lot of details, so take it as an indication of where you'll eventually get to

In special relativity, the energy of a particle of mass $m$ is dictated by its velocity $v$ as $E=\gamma m c^2$ where $\gamma=1/\sqrt{1-v^2/c^2}$

If you expand that expression in powers of $v/c$, you get $E=mc^2+\frac{1}{2}mv^2+\cdots$

so for small velocities the total energy of a particle is approximately its rest energy plus its classical kinetic energy

Similarly, the relativistic momentum of a particle is given by $p=\gamma m v$

and for small $v$ one has $\gamma\approx 1$, so $p\approx mv$ which is just the classical momentum

So the classical definitions of energy and momentum get modified to relativistic forms $E=\gamma m c^2$, $p=\gamma mv$

You'll certainly see that at some point, but take it as writ for the moment

enumaris

got one more interview coming up, hopefully I don't bomb that one :D

Semiclassical

There are some interesting relations between these two expressions. One is that $E=p c^2/v$

danielunderwood

@enumaris I stumbled on a personal question in my last interview...though I'm not sure if that was the goal of the question or not

Semiclassical

Which implies in particular that, if you had a particle with velocity $v=c$, then it'd better satisfy $E=pc$

enumaris

17:21

@danielunderwood I just haven't dealt with probability distributions in so long that I stumbled on it -.-

I gave the integral to solve...but did not realize the symmetry in that integral leads to a simple answer

failz

Semiclassical

another relation you can show algebraically is that $E^2-p^2 c^2=\gamma^2 m^2 c^2(c^2-v^2)=(mc^2)^2$

danielunderwood

They gave you an integral to solve on a phone interview?

Semiclassical

Which shows that, if you have $E=pc$, then you'd better have $m=0$

enumaris

You don't need to solve the integral to solve the problem cus of the symmetry

But I only got to the point of "this is the integral to solve..."

Semiclassical

@enumaris so basically some continuous random variable which has a useful symmetry

enumaris

17:23

and then they had to like give me a bunch of guidance in order for me to realize the symmetry

the question was given X and Y are independent gaussian distributed variables with mean 0 and variance 1 what is P(X>3Y)

I gave that you can do a double integral to solve for that with 2 gaussians in it

Semiclassical

ahh. that actually showed up on MSE lately

not with X>3Y specifically

but in the form Y>aX+b

enumaris

the answer is 1/2 by symmetry lol

somehow I didn't realize that

Semiclassical

yeah

enumaris

No excuses, that was a fail on my part

John Rennie

@enumaris I wouldn't worry too much about it. Interviewers are normally interested in whether you know what you are talking about rather than in a specific problem.

Semiclassical

17:25

3Y has mean zero, so you're dealing with two symmetric r.v.s

enumaris

you can also realize the answer by looking at X-3Y>0 and noticing that nowhere on the left hand side does the joint probability's mean move away from 0

Semiclassical

so you should expect to see X>3Y as often as X<3Y

enumaris

yep

that's another way to look at it

Semiclassical

That said, the way you'd do the integral is kinda funny. The only way I know how to do it is to take a partial derivative under the integral sign, compute the integral, and then integrate

4

Q: Integration of the product of pdf & cdf of normal distribution

Denote the pdf of normal distribution as $\phi(x)$ and cdf as $\Phi(x)$. Does anyone know how to calculate $\int \phi(x) \Phi(\frac{x -b}{a}) dx$? Notice that when $a = 1$ and $b = 0$ the answer is $1/2\Phi(x)^2$. Thank you!

(the fact that that's on MO is hilarious to me)

enumaris

but I failed in all 3 ways and only gave "the answer should be $\int_{-\infty}^\infty\int_{3y}^\infty e^{-x^2} dx e^{-y^2} dy$

Semiclassical

17:27

yeah

enumaris

if you look at what that integral is though

Semiclassical

Actually, if you interpret that geometrically

enumaris

you realize it's just integrating half the plane

Semiclassical

right

enumaris

which leads to another way to realize it's 1/2

Semiclassical

17:28

So that gives another solution

enumaris

so

I failed 4 different ways

lol

Semiclassical

the fact that you can still compute it if you've got $X-Y>a$ is nice

enumaris

anyways

fair play

Semiclassical

fun fact, though: Mathematica can't actually compute $\int_{-\infty}^\infty \phi(x)\Phi(ax+b)\,dx$

even if I tell it to assume $a>0$ and $b$ real

enumaris

I mean, after the dude guided me to the solution

it feels like I was just being extra dumb

XD

Semiclassical

17:31

lol, it happens

lol, yeah: Integrate[ CDF[NormalDistribution[0, 1], x] PDF[NormalDistribution[0, 1], x + 1], {x, -\[Infinity], \[Infinity]}] fails to give a closed-form solution

but the closed-form is just $\frac12(1+\text{erf}(1/2))$

So that's a touch pathetic on the part of mathematica

enumaris

oh wellz, can't help it now, no need to dwell on the past

I got other stuff coming up anyways :D

danielunderwood

17:57

How did you guys learn probability/stats? About all I've had was what I needed to analyze certain bits of data

vzn

@enumaris was it a datascience job? seems rather advanced or somewhat unrelated for that

Slereah

Thank you for your interest. The problem with my thesis proposal is that it requires a good knowledge of some theoretical courses that you may have not taken: advanced QFT, supersymmetry and supergravity and introduction to string theory (besides particles physics of the Standard Model end Beyond, GR and cosmology). If you think you can compete for a fellowship in EDPIF, you can send me a detailed CV with grades of the last years and a list of corses that you have taken.

Dang it

vzn

@Slereah being a phd is very hard, all the cartoons confirm it :o o_O :|

Captain Bohemian

@Slereah who has chance to take these courses? I have never seen any of them taught in either of my undergraduate or graduate school. I even have never taken general relativity. It was not taught throughout my registered academic years. I consider those subjects are for people who research on them to self-study.

Slereah

@CaptainBohemian The students of his master, I suppose

Semiclassical

18:07

I could see advanced QFT/SUSY/SUGRA at certain graduate institutions

albeit more at the research level than the lecture level

ACuriousMind

@Semiclassical advanced QFT/SUGRA/string theory are ordinary master degree lectures in Heidelberg

Semiclassical

whistle

Captain Bohemian

Also, those subjects cover very wide scopes. Courses of the same name, e.g. quantum field theory, taught by different professors may vary greatly.

danielunderwood

Yeah I don't think my school had any of those unless they were special session courses. Though that wasn't really our focus either

Captain Bohemian

I don't particularly like to go to lectures because I found most of my professors in my garduate school didn't teach lectures eaily to follow.

either because their accents are strange, or they presume students have read the textbooks before their courses, or they just can't express science in an understandable ways.

so I didn't usually go to or follow lectures.

danielunderwood

18:22

@enumaris Doing a NN with complex weights wouldn't really do anything other than create extra parameters, right?

Also I didn't realize that the DL specialization had its assignments in jupyter notebooks...that's a lot nicer than the other courses I've seen

Captain Bohemian

my graduate had the course particle physics, but I didn't take it, because I prefer to take quantum mechanics.

danielunderwood

You weren't required to take QM either way?

Captain Bohemian

@danielunderwood not absolutely required to get MSc. But absolutely required for PhD.

We MSc only need to choose two of those fundamental courses.

so if you don't like quantum mechanics, you can choose to take classical mechanics and statistical mechanics or classical electrodynamics.

but by Chance I like quantum mechanics very much, so I prefer to take it than particle physics.

danielunderwood

Ahh I see. MS is pretty rare here as far as I know, so most people take all the fundamentals then electives that go along with their research I think

I am curious how undergrad compares here with outside of the US though

Captain Bohemian

I actually took course based on my own interests rather than research need because my advisor never gave advice about what courses are helpful for research.

danielunderwood

18:38

Yeah I didn't either. But I think I may have taken all my electives by the time I was doing research. I kind of regret not taking extra electives though

Especially since one of my electives was an astro course that I just hated

Captain Bohemian

I afterwrds found quantum mechanics is not used at all in my research.

@danielunderwood not that much time in MSc to take all electives because you need a lot of time to do research.

danielunderwood

Well mine was undergrad, so more courses and less research. But still not a ton of time to take electives with all the non-physics courses that were required

Slereah

Reading all that GR was a mistake really

Nobody's selling GR thesis since the 80's

It's all the QFT and strings these days

Captain Bohemian

but in undergraduate school, I did have taken all electives I can take because 4 years without research is really too long; if not taking courses there is too much free time.

danielunderwood

GR is great though

Slereah

18:44

it is

danielunderwood

And we had a pretty big focus on undergrad research, so I was involved in it most of my 4 years. I thought it was common, but looking back I can only think of a handful of classmates that did the same

Slereah

I really should be brushing up on my QFT really

Captain Bohemian

I was not so dedicated to non-physics courses during my undergraduate studies.

Slereah

I get the feeling it will carry me more for finding a thesis

Time to get Peskin and Schwarze out

danielunderwood

I tried starting Peskin once...I did not get very far. Should probably try it out again

Captain Bohemian

18:47

@danielunderwood I did consider dong research in undergraduate but my undergraduate school didn't have professors conducting research of my interests; most of their research are very applied rather than theoretical.

danielunderwood

@CaptainBohemian we were required to take a good many, but I don't think it's like that outside of the US from what I hear. I wasn't a huge fan, though there was a history course that I enjoyed a good deal

enumaris

@vzn basically data science. They call it "algorithm developer" I guess.

@danielunderwood you'd have complex activations, at the end of the day you would have to get those complex activations and translate it to probabilities or some real number

I think how the complex weight would affect your outputs depend on the way you translate the activations to your desired output

but probably the most natural way will simply lead to you having effectively 2n weights

Captain Bohemian

@danielunderwood we also had history course; I even flunk it once. I don't like history; I prefer geography.

enumaris

I havne't worked in that though so I can't say for sure

welp that last interview didn't seem to go that well lol

not a good day for interviews for me I guess

I asked him why Scala and he attacked python as an inferior language because of its dynamic allocation of memory...so...I guess...that's a no XD

danielunderwood

That would tell me all I needed to know about that job :D

Though every language is going to be allocating memory dynamically if you want to be doing something useful

enumaris

18:55

he said basically Python is not production grade and he had to put that out there

and he needs to know the sizes of his arrays at compile time not at run time

(don't know why)

danielunderwood

I suppose that could be the case if you're doing something embedded with fairly limited resources

enumaris

I asked him what about something like DL4J that ports python based DL frameworks over to Java

and he said the core is still python so it's unacceptable

so, the only thing I can gather from that, is that he wants to manually write the models that he's using

unless there's DL frameworks in Scala that I'm unfamiliar with

which could definitely be a case I guess

but he was pretty critical of tensorflow as well as "super slow"...I don't know...

danielunderwood

Guy sounds like a loon to me

Been drinking too much JVM kool aid

enumaris

lol

I'm not sure DL4J actually just "ports" python code over to Java tho...I think they actually implemented their library in Java...they just have a way to like...load weights or something? I don't know the backend that well tbh...

danielunderwood

Should've asked why they aren't using plain C and CUDA if they're that concerned about performance

Though evidently there's an LLVM backend(?) in development for scala

enumaris

19:00

yeah I dunno lol

maybe just write in assembly

leave any semblence of a compiler/interpreter out of it

Sir Cumference

Howdy

Physics Meta

0

Q: What does conceptual mean?

I've told to come here by another user after my question was put on hold a few days ago for being "off-topic". What on Earth does "conceptual" mean? It's a word I see floating around a lot on this site and it's never explained. Ever. I asked the aforementioned user what it meant and they directe...

discussion scope questions homework-and-exercises

danielunderwood

Or maybe he had a dark history with segfaults

enumaris

he said his background was in R (???) and that he isn't a coder by trade

???

danielunderwood

Though I suppose compile-time allocations may all end up on the stack and be a bit faster (not to mention compiler-optimizable)...that's a good case of premature optimization if I've ever seen it though

enumaris

19:06

whatever, I'm not always gonna find the ideal role for every interview anyways XD

danielunderwood

Ahh then I'd go with the kool aid explanation if that's his background

enumaris

he also said there's little collaboration at that company

you're expected to do your own work

which is another red flag for me

esp for a small start up

danielunderwood

Yeah that would be a big red flag. There's a big difference between "needing to be able to do things independently" and "little collaboration"

enumaris

he mentioned like he could ask an expert for help if he didn't know a particular thing

danielunderwood

I'm in one of the "little collaboration" environments...not so great

enumaris

19:08

but mostly you're gonna be doing your own work

vzn

@enumaris there are not many (serious) data scientists who would devalue python these days. its quite possible his work was doable with python but he had some kind of bias. on other hand, its not unusual for different environements to have different tools.

enumaris

I don't have a problem with people wanting to use something non-python

it's just his justification for it caught me a bit off-guard

danielunderwood

There's a difference in not using it and attacking it as inferior

vzn

sometimes people reject tools they havent used personally on spurious grounds etc

enumaris

in terms of demanding static memory allocation as his root reason...

danielunderwood

19:17

Something something bad worker blames his tools

vzn

bottom line, a very large amount of data science is being done on python now. and it seems to be increasing substantially. also it can be coupled with higher compute resources eg GPUs. so his attitude is basically backward. yeah you run into all kinds during job searches. not every job offered is desirable...! ps have not done a lot of python myself but can see a wave when inundated by it!

enumaris

if the job is like "write deep learning models from scratch, without using anything like tensorflow" then that's a no for me dawg

vzn

lol it sounds unlikely

enumaris

I wrote my own feed forward neural network using numpy

it was a pain in the ass

if I can't even use numpy array indexing and the like...forget about it

vzn

yeah it is not very common one needs to implement the basics anymore.

danielunderwood

19:21

Even if you felt the need to do production with something else, I don't see how you could really build models without something like python to prototype in

And numpy indexing can take a bit to get used to

enumaris

You can write a convolution neural net from scratch

with like just arrays

I can see how it could be done

vzn

python can be used for production. have no idea who claims otherwise...

enumaris

but I sure as hell don't want to be the guy to do it

the backpropagation...lawl...

vzn

there is some (historic) bias that dynamic languages (in contrast to compiled ones) arent suitable for production but think its a fading concept. the distinction is blurry anyway. eg java crosses the boundaries etc

enumaris

you'd be writing numerical differnential equation solvers at the end of the day

and at the end of the day that sure as heck not what I'm interested in

danielunderwood

19:23

Yeah I have a variety of things using python in production. There are a few reasons not to use it, but those are kind of edge cases

enumaris

is dynamic memory allocation one of them? I have basically no insight into that base level of programming.

I generally don't care about what a compile or an interpreter does, I only care about my inputs and outputs

vzn

someone might argue that some apps push the garbage collection of dynamic memory too hard, but suspect thats more to be seen as bad coding... have worked with very memory intensive apps in java and the (highly tuned!) gc is almost never a factor.

enumaris

But say you built and trained a DL model in python using tensorflow

you have all the weights

and you port it over to java using DL4J

danielunderwood

I haven't run into anything where dynamic memory allocation was a problem. There are people that work on platforms with crazy resource restrictions, but I've worked on fairly limited microcontrollers with dynamic allocation being fine

enumaris

how is that all of a sudden not production grade?

I don't really get it =/

vzn

19:26

@enumaris the guy may not even know what youre talking about. theres more than 1 way to skin a cat™... dont assume everyone knows what theyre talking about o_O

enumaris

DL4J has a scala extension? port? w/e as well

you can port DL models directly over to scala using DL4J as well

danielunderwood

I've heard some people call python not production grade because it's slow though that's rarely a big issue and especially not an issue if you're just wrapping C calls. Some people may say that because python doesn't require a given structure, so it's a lot easier to write sloppier code as project size increases

Scala should be able to import java libs, but I'd imagine using them can be a pain due to the difference in Java and Scala

vzn

its possible to mess up the code in any language...

@danielunderwood sometimes really bad workers blame others tools :P

enumaris

In any case, probably if the language of the code is a huge driving force, either they have to have software engineering support to code in Scala, or probably it's not the job for me lol. I'm happy to learn Scala for a job, but it ain't gonna be like "implement a DL framework from scratch in Scala". XD

vzn

scala is a rising language also, dont know a lot about it myself, wouldnt dis it. seems to have some niches.

danielunderwood

19:33

I will say that python can have issues in that it can run into issues during execution that would be caught by the compiler in other languages. Solid testing should help that though

I think that scala is sold to people as easy to parallelize, which is indeed the case when it comes to functional code. But it's not like it's difficult to write parallel or distributed code in python if you roughly know what you're doing

enumaris

yeah I have no problems with any language

I am willing to learn a language to work within a company

but my ideal role is just not implementing DL frameworks from scratch or something like that...moving pointers around and the like...

figuring out if backpropagation is doing what it's supposed to be doing...

debugging backpropagation is painful to the max

danielunderwood

I still don't think I fully understand backprop...though I haven't needed to so far

enumaris

it's just the chain rule bro

applied iteratively

literally that's it...

but debugging it is P.A.I.N.

danielunderwood

Ahh that isn't bad...the way Ng explains it seems a bit convoluted to me. Though I think he takes the assumption that people don't know calculus

I did have to explain the chain rule to one of my classmates once...it was far too late in our courses to have been doing that

enumaris

You could use a numerical estimate in place of backprop, like for each weight $W$ just use $\Delta J/\Delta W$ as your gradient...it's just that doing that for every single weight is computationally inefficient. It's better to calculate $\Delta J/\Delta W$ using the chain rule

XD

You end up with something like $J(A(B(C(...(W)))$ and then you apply the chain rule to that and the expression like $\partial J/\partial A \cdot...\partial A/\partial B...$

is easier to evaluate computationally since you don't have to make a full forward pass to get the answer, you can get it from matrix multiplication

danielunderwood

19:44

That's no problem then. I remember him saying something about errors on the output that was really confusing. Should have just worked it out myself

enumaris

yeah the first time he presents it, he's assuming a sigmoid activation

danielunderwood

Although that's my strategy going through this DL series. Just taking minimal notes and working out the derived results when I need them

enumaris

and for a sigmoid activation the partial deriviative term that matters turns out to be like the actual different between the predicted and real values

that was confusing to me at first for his course as well

danielunderwood

Ahh I see

enumaris

it's a constraint of him not wanting to go into calculus

danielunderwood

19:46

I feel like anyone in DL should have a knowledge of calc, but maybe that's just me

Or data science/stats/whatever

enumaris

well you could certainly just use tensorflow

it would remove backpropagation from consideration and voila you don't need the chain rule anymore lol

but yeah probably having a quantitative background would help a lot

danielunderwood

But then you could still choose a stupid loss function and have no idea what's going on

enumaris

probably...

danielunderwood

Though I suppose that could be solved with "make sure the loss only has a single extremum"

enumaris

^that is highly non trivial

danielunderwood

19:48

Yeah I honestly have no idea how to prove that for a given function

But it rules a lot of stuff out at least

"Sine error"

Enzo

Hi there

I'm trying to develop a project in such a way as to prevent traffic accidents using Electromagnetism.

enumaris

also you gotta reword it

Enzo

However, I dont have a deep knowledge in this related field to get my questions answered concerning charging a stuff to make them apply repulsive force to push themselves apart. For instance, suppose that two cars are accelerating and about to get crashed. There will be two positive/negative charges

enumaris

"single extremum" is not what you want I think XD

maybe...it is...

Enzo

My main goal is to charge a stuff either positive or negative to make them apply repulsive force as stated earlier

enumaris

19:55

everywhere convex function with only one global minimum?

I think you can use functions a bit more general than that tho

@Enzo you'd probably get a lightning bolt somewhere before you got enough charge to apply a significant repulsive force

but that's just my guess

Enzo

@enumaris What did you mean by lightning bolt?

enumaris

there's only so much potentially difference that air can sustain before it becomes conductive and you get a spark

lightning bolt :D

Enzo

I'm still unable to get it lol.

Is this project also possible when we think theoretically?

danielunderwood

@enumaris yeah that's probably a better description. I was thinking extremum since you could have a global maximum and just negate the function, but that's extra steps. I thought one global minimum would imply everywhere convex though

It's very difficult to sustain any net charge on an object. If you give yourself a charge by shuffling on the carpet, it discharges as soon as you touch anything that it can go to

Enzo

@danielunderwood but what else can I do?

danielunderwood

20:16

I don't know that there's too much you can do. If you managed to give a car a charge, it would want to escape. You also can't magically bring this charge out of nowhere, so you'd have to just redistribute the charge that's already on the car. That's going to make any repulsion you have be an attraction elsewhere. If I wanted to make a car stop to prevent an accident, I'd use sensors and a mechanical means of stopping

You could perhaps cook up something with magnets that's less volatile, but I still don't see it working well in the real world

Though I'll also say that I'm no expert in using EM to achieve practical things

You could probably build a toy model with a magnet to see how it works, but doing the same with trying to build a charge could be quite dangerous and I'd suggest against it

bolbteppa

20:50

Physics is cool

Theoretical variety cooler

Slereah

The problem being that you also make theoretical money

enumaris

I make lots of theoretical money

Transcript for