The Big Programming Thread - Page 1002

Prev 1 1000 1001 1002 1003 1004 1032 Next

Thread Rules
1. This is not a "do my homework for me" thread. If you have specific questions, ask, but don't post an assignment or homework problem and expect an exact solution.
2. No recruiting for your cockamamie projects (you won't replace facebook with 3 dudes you found on the internet and $20)
3. If you can't articulate why a language is bad, don't start slinging shit about it. Just remember that nothing is worse than making CSS IE6 compatible.
4. Use [code] tags to format code blocks.

Thaniri

1264 Posts

March 02 2019 17:19 GMT

#20021

I guess the drunken version of me was more abrasive than he needed to be. My bad.

I was more viewing both sides of the coin. The thread really did look like one dude saying college is useless and one dude saying cs logic is essential.

I'll bow out because it's not a hill I am willing to die on. The smartest guy I ever worked with doesn't have a highschool diploma. However he is 1 guy and all of my other smart coworkers have degrees. Statistics would bear out that the degree is useful in most people's eyes in that regard.

Deleted User 3420

24492 Posts

March 02 2019 18:02 GMT

#20022

Yep of course, a degree is not needed, especially in this age of the internet. Brilliant driven people can learn most anything on their own these days. But people like me do much better with some kind of external stimulus driving them. Plus degrees are a good way for an employer to filter out nonsense (assuming they are from a reputable school).

Deleted User 3420

24492 Posts

March 03 2019 18:05 GMT

#20023

Hey so I was looking at images of feedforward neural network and I see two different things come up:

#1:

http://web.utk.edu/~wfeng1/spark/_images/fnn.jpg

#2:

https://c7.uihere.com/files/261/538/457/artificial-neural-network-deep-learning-convolutional-neural-network-backpropagation-feedforward-neural-network-neural.jpg

In image 1, earlier outputs are feeding into the bias units of the later layers. In image 2, this is not the case.
I am guessing that image #2 is incorrect? Bias nodes are just a constant value for purposes of making linear algebra work, right?

SC-Shield

Bulgaria859 Posts

March 05 2019 22:23 GMT

#20024

Wow, I didn't expect re-reading a book to be so productive. I guess I didn't pay enough attention the first time, so I missed syntax like:

C++11


<return type> FunctionName() &;
<return type> FunctionName() const&;
<return type> FunctionName() &&;

Also, a few other C++ details. I guess during first reading I was focused on learning new features, and now I can appreciate all the small details.

Edit: It's funny how the C++ committee forgot to add std::cbegin(), std::cend(), std::crbegin() and std::crend() in C++11.

berated-

United States1134 Posts

March 06 2019 03:23 GMT

#20025

On March 02 2019 14:14 Thaniri wrote:
What forums/blogs/websites/books do you guys read for system design topics?

I just realized that when I have a random question on how best to design a system, I don't actually have a place to post and discuss or search for an answer.

For example, the one that just came to mind:

How do people loadbalance web application front ends in the cloud? Traditionally, we put a webserver in front of of a bunch of webapps and just scale the number of webapps as required. However, with the magic of auto scaling groups, in order to do that, you need to make a config change to your balancer manager every time a new web app server gets brought up or deleted.

So due to the magic of the cloud we can do things like ALBs in AWS. But does that mean we no longer use nginx/apache? All the configs like reverse proxy rules, ip blocking configs, and SSL certs can't be expressed in AWS ELB rules. And doing something like a letsencrypt cert renewal would be hard for an ELB. You'd need to somehow do the renewal on a client, run a terraform/cloudformation template that takes the new cert from the client and put it on the server.

The most idiot proof thing I can think of is to put a webserver onto every server with a webapp running. Put that server type into an autoscaling group, and delegate ALL loadbalancing to the ELB. That way any complex webserver configs are retained. The obvious issue with this is that you will end up with a silly number of webservers in a given deployment.

Also a valid answer to this would be "go learn kubernetes for amazing autoscaling and self healing you caveman" but I'm a luddite who still uses bare metal. And not enough bare metal to make a borg cluster

When you find an answer of where to go, please let everyone know, including me. Some people pay big $$$ for that information.

Your proposal is pretty much what we are doing. We run apache servers on every node of our webapps to retain the things like you are talking about. We run java webapps, and the apache along side of it. The overhead is basically negligible, so yeah.. you have an extra process on all your nodes, but you were going to have that number of nodes anyways. Load balancing and TLS termination is done at the LB level. I'm more on the dev side, but, on prem cert management is a pain in the ass. I've heard doing so in amazon is super simple. Even if you had let's encrypt going, managing certs on a top level LB is easier than doing so on each apache.

That being said... if what you are doing is currently not working - then imho, yes, kubernetes is the future. The learning curve is massive, building it yourself is frowned upon (by some) ...but the abstractions and abilities are incredible. If you're already in the cloud, then using someone to manage the cluster for you is ideal.

sabas123

Netherlands3122 Posts

March 07 2019 10:47 GMT

#20026

Hey travis, how did your quest for solving that NP problem go?

Neneu

Norway492 Posts

March 07 2019 11:28 GMT

#20027

On March 04 2019 03:05 travis wrote:
Hey so I was looking at images of feedforward neural network and I see two different things come up:

#1:

http://web.utk.edu/~wfeng1/spark/_images/fnn.jpg

#2:

https://c7.uihere.com/files/261/538/457/artificial-neural-network-deep-learning-convolutional-neural-network-backpropagation-feedforward-neural-network-neural.jpg

In image 1, earlier outputs are feeding into the bias units of the later layers. In image 2, this is not the case.
I am guessing that image #2 is incorrect? Bias nodes are just a constant value for purposes of making linear algebra work, right?

Slightly depends on the problem you are supposed to solve, but the general idea is that the main function of a bias is to provide every node with a constant value which can be trained. This can be achieved by a single bias node with connections to N nodes or with N bias nodes each with a single connection, the result should generally be the same. So even if in image two they only have one bias node (i = 0) per layer, it should work as well as in image one. But yeah, not a big fan of how they did the diagram.

ZenithM

France15952 Posts

March 07 2019 12:18 GMT

#20028

On March 07 2019 19:47 sabas123 wrote:
Hey travis, how did your quest for solving that NP problem go?

That's the topic that makes me single-handedly stop checking on this thread for 3 months.

Deleted User 3420

24492 Posts

March 07 2019 12:52 GMT

#20029

zenith that doesn't even make any sense. why don't you just go away forever instead of only 3 months

sabas:

I ended up learning a lot, particularly about hamiltonian cycles. Made some interesting heuristics that work some of the time even for decently large n. I want to keep investigating but I have to do school stuff and it's just so time consuming.... like incredibly time consuming to do any detailed exploration of the behavior of these graphs and how they evolve with more nodes or edges added.

Manit0u

Poland17774 Posts

March 07 2019 13:08 GMT

#20030

@travis: have you ever tried stuff like that?

https://github.com/koudelka/visualixir

ZenithM

France15952 Posts

March 07 2019 13:41 GMT

#20031

Come on Trav, you post like 10 times a day on this thread (again, on subjects ranging from high school trigonometry to millennium prize problems), you can handle one contrarian comment about your P=NP solution once in a while.
Granted this time it wasn't your fault, you were prompted to give us an update.

WarSame

Canada1950 Posts

March 08 2019 04:53 GMT

#20032

As long as he doesn't dominate the thread with his P=NP talk(which he doesn't) then acting so negatively about someone's passion is being a dick.

ZenithM

France15952 Posts

March 08 2019 14:32 GMT

#20033

I agree that I'm unfairly being a dick to him (I apologized 3 months ago but I couldn't resist doing it again), but I felt, perhaps wrongly, that there needed somewhere in the sea of positiveness and support to be some kind of devil's advocate voice for how ridiculously pointless it is to work on P vs NP. I remember everyone was being supportive of him at the time, and I don't think it helps travis in his progression as a computer scientist (myself I told him back then that it would be more fruitful for him to focus on more productive endeavors, which he seems to be doing now).
Any CS professor would have told him to focus on something else than P=NP (especially this particular side of the P vs NP question...).

It's really a testament to this thread's welcoming atmosphere that there is only one guy making fun of another for working on P vs NP. It's impressive and I'm a bit ashamed to be that guy now but I also don't dominate the thread with snarky posts either. Anyway, I won't post again on that topic if it comes up, even if it's 3 months from now

. Consider me self-banned from doing that.

solidbebe

Netherlands4921 Posts

March 08 2019 16:02 GMT

#20034

Everyone was being supportive? Don't know which thread you've been reading.

enigmaticcam

United States280 Posts

March 10 2019 02:44 GMT

#20035

What’s wrong with someone doing what they want in their own free time?

Excludos

Norway8263 Posts

March 10 2019 03:10 GMT

#20036

You guys have spare time..? As programmers?! Unheard of!

Chocolate

United States2350 Posts

March 10 2019 06:43 GMT

#20037

I prefer not to think of a FFNN with "bias units". FFNN are essentially multilayer multi-perception networks. In the perceptron model each component of a layer is a perceptron consisting of a weight for each input and a bias.

So both these pictures are just different ways of representing bias. I actually think the first picture is the worse because it separates out the bias from the nodes. The second picture more clearly conveys that each node has a bias value.

Chocolate

United States2350 Posts

March 10 2019 07:01 GMT

#20038

Does anybody here do distributed computing? I work on it in my day job and want to know what the next new hot thing to study is - like Spark before it was Spark. My manager seems to think it's Presto but Presto honestly seems kind of shitty in terms of performance. Dask is a little old but looks like adoption is growing rapidly and it seems to be popular in quant firms.

Neshapotamus

United States163 Posts

March 13 2019 05:21 GMT

#20039

The geometric interpretation of the bias is the constant value that lets you shift your graph vertically. Technically, this shouldn't be multiplied by any value in the previous layers unless you want to have the property that the function goes through the origin.

ex: y=w1*z1+w2*z2+..w_n*z_n+b

SC-Shield

Bulgaria859 Posts

March 13 2019 18:34 GMT

#20040

Could you please recommend a few nice books about Machine Learning? If they're about Reinforcement Learning, then it will be even better.

Prev 1 1000 1001 1002 1003 1004 1032 Next

Please or register to reply.

The Big Programming Thread - Page 1002

Completed

Ongoing

Upcoming