Members in Shoutbox
None.

Shoutbox Search
Search for:


Shoutbox Commands
/w [name] > Whisper
/r > Reply to last whisper
/me > Marks as action

Shoutbox Information
Moderators may delete any and all shouts at will.
Global Shoutbox
Please log in to shout.
Pages: < 1 « 1733 1734 1735 1736 1737 … 3578 >

[2016-5-02. : 6:16 pm]
Roy -- Vrael
Vrael shouted: that's because you're a static character in the novel of my life and don't learn or change at all
Nonsense; I've changed. I used to be Roy the Regular, remember?
[2016-5-02. : 6:15 pm]
KrayZee -- It's an app on iPhone
[2016-5-02. : 6:15 pm]
Vrael -- then how did he hear you needed a Lyft?
[2016-5-02. : 6:14 pm]
KrayZee -- Interesting, I requested a Lyft ride and the driver is deaf.
[2016-5-02. : 6:14 pm]
Vrael -- Tikels on the other hand, has a chance to grow
[2016-5-02. : 6:14 pm]
Roy -- It's solid advice that almost nobody will follow.
[2016-5-02. : 6:13 pm]
Vrael -- you've just always been, and always will be, Roy
[2016-5-02. : 6:13 pm]
Vrael -- that's because you're a static character in the novel of my life and don't learn or change at all
[2016-5-02. : 6:13 pm]
Roy -- I'd repeat my same procrastination mistakes if I were given the chance all over again.
[2016-5-02. : 6:12 pm]
Vrael -- Roy
Roy shouted: Vrael Haven't you ever been a normal student?
I'm helping him learn from my mistakes
[2016-5-02. : 6:11 pm]
Roy -- Vrael
Vrael shouted: TiKels how come you're only working on your project the day its due? :P
Haven't you ever been a normal student?
[2016-5-02. : 6:11 pm]
Vrael -- but they're completely not useful because no real-world application ever has all the sampling and feature extraction and formatting and shit already done
[2016-5-02. : 6:11 pm]
Vrael -- well thats not entirely true, if you've had a team of engineers already come in and sanitize the data and do feature extraction and format everything nicely for you in some standardized database (basically all the real work) then there are over 9000 software packages with GUIs that pretend like they're useful
[2016-5-02. : 6:09 pm]
Vrael -- TiKels
TiKels shouted: Also I should mention that this is not some kind of gui, all the code is just Matrix math and standard functions. That's what makes it even more frustrating
gui's are for squares anyway, if you're using a gui while doing ML you're probably doing something wrong
[2016-5-02. : 6:09 pm]
Vrael -- though I don't think he covers neural networks
[2016-5-02. : 6:08 pm]
Vrael -- theres also plenty of free Coursera courses, my favorite is just Machine Learning by Andrew Ng
[2016-5-02. : 6:08 pm]
Vrael -- you can find it free online
[2016-5-02. : 6:08 pm]
Vrael -- Yasser Abu Mostafa I think is his full name
[2016-5-02. : 6:07 pm]
Vrael -- Clokr_
Clokr_ shouted: Vrael, it does sound interesting. Do you have any good references to get started on this?
there's a great free CalTech course which includes linreg, logreg, SVM's and Neural Networks by prof. Abu Mostafa that I highly recommend
[2016-5-02. : 6:07 pm]
Vrael -- Clokr_
Clokr_ shouted: Vrael, it does sound interesting. Do you have any good references to get started on this?
there'
[2016-5-02. : 6:06 pm]
Vrael -- TiKels
TiKels shouted: Great. You all come out of the woodwork an hour before my project is due
how come you're only working on your project the day its due? :P
[2016-5-02. : 6:04 pm]
jjf28 -- http://www.staredit.net/352320/ in fact, your current string section only works cause sub-string recycling :O it'll take me a while to get that coded up
[2016-5-02. : 5:35 pm]
jjf28 -- 64254 characters divided into 1032+ strings is way too much
[2016-5-02. : 5:30 pm]
jjf28 -- you're almost definitely at the string limit zzsnake
[2016-5-02. : 4:51 pm]
TiKels -- I might continue learning about this stuff. I find it really cool
[2016-5-02. : 4:48 pm]
TiKels -- Or well, the day of
[2016-5-02. : 4:48 pm]
TiKels -- Great. You all come out of the woodwork an hour before my project is due
[2016-5-02. : 4:47 pm]
dumbducky -- Oh shit are we taking neural networks?
[2016-5-02. : 4:38 pm]
Clokr_ -- Vrael, it does sound interesting. Do you have any good references to get started on this?
[2016-5-02. : 4:27 pm]
TiKels -- I think the most complicated thing in there is matrix inverse
[2016-5-02. : 4:26 pm]
TiKels -- It's like 50 lines of code or less
[2016-5-02. : 4:26 pm]
TiKels -- Also I should mention that this is not some kind of gui, all the code is just Matrix math and standard functions. That's what makes it even more frustrating
[2016-5-02. : 3:46 pm]
TiKels -- What kind of research do you do?
[2016-5-02. : 3:41 pm]
NudeRaider -- dark :) http://i.imgur.com/VKS90o2.jpg literally.
[2016-5-02. : 3:36 pm]
Vrael -- I'm a research scientist [BUZZWORD ALERT]DATA SCIENTIST[/BUZZWORD ALERT}
[2016-5-02. : 3:34 pm]
Vrael -- I build ML models every day
[2016-5-02. : 3:34 pm]
Vrael -- this shit is my job
[2016-5-02. : 3:34 pm]
Vrael -- well that doesnt sound degenerate
[2016-5-02. : 3:34 pm]
TiKels -- Good for you
[2016-5-02. : 3:34 pm]
TiKels -- I'm impressed you're this well versed vrael
[2016-5-02. : 3:33 pm]
TiKels -- I know I can alter the number of layers but that's a bit outside the scope of my knowledge. The neuron count is pretty simple though
[2016-5-02. : 3:32 pm]
TiKels -- I was doing 10 neurons with one hidden layer.
[2016-5-02. : 3:32 pm]
Vrael -- if the other setup was specifically tuned to that data you could have like, a single node NN or something degenerate like that which isn't working
[2016-5-02. : 3:31 pm]
Vrael -- if you have parameters for those
[2016-5-02. : 3:31 pm]
Vrael -- well, maybe try a less restrictive set of hidden layer depth/width
[2016-5-02. : 3:30 pm]
Vrael -- ok
[2016-5-02. : 3:30 pm]
TiKels -- No no, I trained it again. I meant the same method of training
[2016-5-02. : 3:30 pm]
Vrael -- you can't train on a data set of shoe sizes and expect to do facial recognition
[2016-5-02. : 3:29 pm]
Vrael -- you need to train it again on the X's and Y's for each data set
[2016-5-02. : 3:29 pm]
Vrael -- ok well, don't use the exact same network
[2016-5-02. : 3:29 pm]
TiKels -- Using the exact same network it was able to perfectly approximate a different set of data, but if I switch the data to my case it just looks awful
[2016-5-02. : 3:29 pm]
Vrael -- thats probably the best, though I also support your decision to go with a simpler model lol
[2016-5-02. : 3:28 pm]
TiKels -- That's just the thing. I don't know enough about the method to actually tell you what's going wrong. I'm gonna ask the prof about it
[2016-5-02. : 3:26 pm]
TiKels -- You mean my training data? It's handed to me. I don't get to generate it
[2016-5-02. : 3:26 pm]
Vrael -- a NN should have no problem generalizing something that simple
[2016-5-02. : 3:26 pm]
Vrael -- maybe you're generating x and y wrong
[2016-5-02. : 3:26 pm]
Vrael -- well, like I said, look at your feature set
[2016-5-02. : 3:25 pm]
TiKels -- Because it actually was able to approximate the function
[2016-5-02. : 3:25 pm]
TiKels -- I eneed up swithing to a method using radial feedback with gaussIan activation functions
[2016-5-02. : 3:23 pm]
TiKels -- There was a set of x and y values that looked similar to a sine wave. I need to make a neutral network approximate this function and then determine its effectiveness in extrapolation and interpolation. Then, perform sensitivity analysis as to how well it converges when changing initial conditions (eg learning rate, iteration count, etc)
[2016-5-02. : 3:21 pm]
Vrael -- but what's your question?
[2016-5-02. : 3:20 pm]
TiKels -- Vrael
Vrael shouted: there's no reason a NN should be limited to only radially symmetric data
i was trying to use a particular neural network structure to do curve fitting. The method is called back propogation. The activation/threshold function was bipolar sigmoid.
[2016-5-02. : 3:12 pm]
Vrael -- so take a good hard look at what your feature set looks like
[2016-5-02. : 3:12 pm]
Vrael -- but a shitty algorithm with great features will outperform the strongest algorithm with shitty features
[2016-5-02. : 3:12 pm]
Vrael -- unless you're trying to do automated feature discovery with 'deep learning' you probably don't even need a neural network and will get equivalent results with a faster algorithm, if training time is an issue for you
[2016-5-02. : 3:11 pm]
Vrael -- however, the point most people miss is that the NN probably is not the problem, most likely it's your feature set
[2016-5-02. : 3:10 pm]
Vrael -- there's no reason a NN should be limited to only radially symmetric data
[2016-5-02. : 3:04 pm]
Vrael -- also to use them properly you di need to know how they work and why they work and all those other important details, or you'll end up like my co-workers who use this shit wrong and build shitty models that don't work
[2016-5-02. : 3:03 pm]
TiKels -- I took a class on AI and we spent 3 weeks on neutral nets.
[2016-5-02. : 3:03 pm]
Vrael -- Sacrieur
Sacrieur shouted: It's a waste of effort on my part to even attempt it, since I'm not going to school for eight years to learn about artificial intelligence and neural networks.
also 1 semester class will probably bring you up to date

Pages: < 1 « 1733 1734 1735 1736 1737 … 3578 >


Members Online: ArjenVeldhuizen459, MarloesKieft554, JochemRietbergen266, WiebeZwartjes491, EelkjeZandstra904, MauritsHellemans569, RosalienMolenaar254, HannekeWiersema947