|
| 30 Aug 2016 12:10 AM |
in 1960-something, neural networks failed to process natural language and correctly process it into a different natural language. They can't tell things like tone or even word structure, implications, or really meaning. Let alone emotion or this buzzword 'toxicity'. It's an absolutely awful way of parsing natural language. It makes *way* more mistakes than what it ever gets right.
what makes it any different 50 years later, a neural network is still a neural network. Especially when all you've done to 'solve' the problem is just throw more processing power at it with moore's law. That changes nothing. Accuracy 76% of the time is a failure, Accuracy 99.999% of the time still has room for error. You've failed until you get to something like 99.99999999999% accuracy, and nothing comes even remotely close to that yet.
it's still a failure and it's still garbage and no one has figured out how to properly implement an AI that works akin to how x86 emulators (or any other IS, for that matter) do for x86 chips. You need structure similar to synapses and you need to emulate the brain, not try to "figure out how to do something like it" because if you don't brute-force it, you're not doing it the same way, hence, causing yourself a lot of pain trying to shoehorn something where it very much does not belong.
The brain is a network of billions of little low-power cores just everywhere that run on something nothing like what we have built today.
You need hardware emulation before you can try software emulation and get it right. millions and millions and millions of ARM cores.
i'd like to say that within 5, maybe 10 years there's going to be another AI crash. Like the last time. And the one before that. And that's happened like 6 times now. Everyone always forgets about it - it __always__ happens.
Thoughts?
tl;dr c-ommu-ity s-ift is written by a bunch of idiots who rather than solve any problem, just go in and do the same thing again that's been done a million times and proven to not work.
please stop using it. |
|
|
| Report Abuse |
|
|
| |
|
| |
|
| |
|
|
| 30 Aug 2016 12:49 AM |
Deep learning, the traditional neural network, is a model that's been proven a million times to be god awful at language parsing, because there is a lot of intricacies to how speech works.
I can say, "The dog went outside." And the machine will know I am talking about a dog and it went outside. Then I can say "Outside was the place the dog went." And the machine will know I am 'talking' about how 'outside went to dog'.
Give it something when it's been trained a specific way outside of how it normally works and it goes bonkers.
that's a fairly incorrect example but it's good enough to illustrate the point.
Or how under the right conditions, a specific word (First one that comes to mind is a synonym of Democrat) can be both 'omg mega SJW offensive' and normal at the same time, under context. Normal deep learning has zero clue about context.
recursive deep learning works along the lines of saying "Outside is here, here is a verb, the subject is there", then re-arrange the words in such a way that it goes back to being "The dog went outside", and be happy with the result.
(again, misleading, but correct enough to illustrate the point.)
of course accuracy comes at a cost, and that is hardware resources. But there's no reason why a company like ROBLOX can't spend more money on servers if they're so passionate about censorship.... servers aren't *RIDICULOUSLY* expensive unless you want 2 ultra mega Xeons and 4 super giga Quadro or Tesla GPUs for your AI stuff. |
|
|
| Report Abuse |
|
|
|
| 01 Sep 2016 06:32 PM |
| the forums look great in 4K |
|
|
| Report Abuse |
|
|
Kayuune
|
  |
| Joined: 08 Jul 2014 |
| Total Posts: 5394 |
|
|
| 01 Sep 2016 06:33 PM |
| What? Who cares about any of this? |
|
|
| Report Abuse |
|
|
| |
|
| |
|
|
| 02 Sep 2016 10:26 PM |
##### ### ### #### ### ### ### ### # ### ## #### # ####
r+://471495989r+://471533497 |
|
|
| Report Abuse |
|
|
| |
|
|
| 02 Sep 2016 10:27 PM |
what a wumbo
scientifically can't touch this https://www.roblox.com/mchammer-item?id=439838249 | https://www.roblox.com/MACHO-MAN-RANDY-SAVAGE-item?id=483161909 |
|
|
| Report Abuse |
|
|
|
| 02 Sep 2016 10:34 PM |
You make no sense, tf
--Life would be tragic if it weren't funny.👻ᵇᵒᵒ ✖ Delete Post |
|
|
| Report Abuse |
|
|
|
| 02 Sep 2016 11:04 PM |
oh boy
Neural networks are stupid because they've been stupid for 50 years. There. Really simple. |
|
|
| Report Abuse |
|
|
|
| 03 Sep 2016 12:15 AM |
| https://forum.roblox.com/Forum/ShowPost.aspx?PostID=197647925 |
|
|
| Report Abuse |
|
|
MrKazz
|
  |
| Joined: 10 Apr 2011 |
| Total Posts: 1212 |
|
|
| 03 Sep 2016 12:25 AM |
@DoAnAileronRoll,
As somebody who used to post regularly in Scripting Helpers, when it used to be called that, I understand why you would be and should be upset about the ridiculousness created from completely ineffective and bothersome auto-censor algorithms. After years of updates that ruined my code and dozens of times being banned because one of my posts was grossly misinterpreted, I came to the realization that the Roblox community doesn't care, the Roblox moderation staff doesn't care, and the Roblox developers don't care. They only care about producing new trendy features that entertain the majority of the underage Roblox target audience. Posts like these really do nothing. If you really wanted to get your point across, I'd recommend the best course of action to be emailing somebody on the Rbx staff. |
|
|
| Report Abuse |
|
|
|
| 03 Sep 2016 12:51 AM |
I see.
Well then, I have some spare hardware lying about here... there's always Ubuntu Server....
|
|
|
| Report Abuse |
|
|