I want to crack Neural Networks

  • 198 Replies
  • 7032 Views
*

LOCKSUIT

  • Emerged from nothing
  • Trusty Member
  • ***********
  • Eve
  • *
  • 1456
  • First it wiggles, then it is rewarded.
    • Enter Lair
Re: I want to crack Neural Networks
« Reply #105 on: April 15, 2018, 04:34:17 pm »
I'm sorry. I really don't know enough to say trees are out of the game.


That's not what I meant. What I meant was, I want to know how a tree can transform a sentence into a different but similar/on-topic sentence.

Ex. the tree transforms "Flower are pretty looking." into "Flowers are a beautiful thing on Earth." Is that possible with trees? Show me. My brain was able to do it. Here's another I'll generate "I believe flowers are a nice-looking thing that have come on Earth from evolution.". "Flowers are certainty not one of the uglier things.".
Emergent

*

ivan.moony

  • Trusty Member
  • *********
  • Terminator
  • *
  • 948
    • Structured Type System
Re: I want to crack Neural Networks
« Reply #106 on: April 15, 2018, 04:41:18 pm »
I'm sorry, Lock, I didn't mean to dispute any of trees or NN for solving a task. I guess I wasn't clear enough. I just wanted to inform you of current top technology known to me for storing natural language sentences.

As for transforming one sentence to another with the same meaning, I'm not aware of any technology or program that does that. But an interesting question, though. If someone figures out how to do it, it could be an interesting base for a compression algorithm. I think it could be done with NN looking backwards, down its synapse tree, but I'm not an expert for NN.
Dream big. The bigger the dream is, the more beautiful place the world becomes.

*

LOCKSUIT

  • Emerged from nothing
  • Trusty Member
  • ***********
  • Eve
  • *
  • 1456
  • First it wiggles, then it is rewarded.
    • Enter Lair
Re: I want to crack Neural Networks
« Reply #107 on: April 15, 2018, 04:48:12 pm »
"Grass loves it when you cut it so keep up down your cutting spree."
"Grass is happy when you cut it so do cut it whenever you can."
"Please cut grass it loves the feeling."
"Cutting grass makes grass feel good."
"Grass will never ever dare cut itself until it changes species."
"The grass wherever it lies can't cut itself without arms because it can't ok grass needs a punk to cut its bosom."
Emergent

*

unreality

  • Starship Trooper
  • *******
  • 277
  • SE,EE,Physicist,Philosopher. Built world's 1st AGI
Re: I want to crack Neural Networks
« Reply #108 on: April 15, 2018, 06:40:20 pm »
[snip] I want to know how a tree can transform a sentence into a different but similar/on-topic sentence.

Ex. the tree transforms "Flower are pretty looking." into "Flowers are a beautiful thing on Earth." Is that possible with trees? Show me. My brain was able to do it. Here's another I'll generate "I believe flowers are a nice-looking thing that have come on Earth from evolution.". "Flowers are certainty not one of the uglier things.".

That level of task is for a developed AGI. That is, an AGI with sufficient learning experience regarding your question. If you're interested I could copy my main tree search function and strip out the code. I'm a ridiculously heavy code commenter lol.

*

LOCKSUIT

  • Emerged from nothing
  • Trusty Member
  • ***********
  • Eve
  • *
  • 1456
  • First it wiggles, then it is rewarded.
    • Enter Lair
Re: I want to crack Neural Networks
« Reply #109 on: April 15, 2018, 07:24:07 pm »
Yes I would like to see that.
Emergent

*

unreality

  • Starship Trooper
  • *******
  • 277
  • SE,EE,Physicist,Philosopher. Built world's 1st AGI
Re: I want to crack Neural Networks
« Reply #110 on: April 16, 2018, 06:51:08 pm »
Yes I would like to see that.
I've been wanting to make a short video about this for sometime. When time permits. Probably this week. Hopefully it will also outline some processor circuit design ideas for tree search. A few decades ago computer gaming relied upon the general CPU to render texture maps. Then came the GPUs, now rendering roughly 370 billion T/second, but it didn't start like that. It seems the STG-2000 was the first GPU, rending just 12 MT/s (0.012 GT/s). Back then I'm sure that was amazing. Poor CPUs. The GPU was also great for neural networking due to matrix math, but it wasn't designed specifically for NN. So google spent tens of millions of dollars making the TPU, now capable of an insane 45 TFLOPs per TPU. One day there will be processors designed specifically for AGI tree search code where the ASIC transistors will contain the tree search and pattern recognition routines, all connected to what I call parallel RAM. That will be the dawn of ASI. :)

*

LOCKSUIT

  • Emerged from nothing
  • Trusty Member
  • ***********
  • Eve
  • *
  • 1456
  • First it wiggles, then it is rewarded.
    • Enter Lair
Re: I want to crack Neural Networks
« Reply #111 on: April 17, 2018, 11:37:04 am »
I'll look up AlphaZero's tree use.

I drew a pic showing why trees are too clear-cut. It's attached.

Here is what i got:
AlphaGo Zero: Uses Residual Network. Doesn't take in human data like AlphaGo and then discovers, rather just self-plays. Starts out with just self-play, backpropagation error rate RL. Monte carlo tree search stabilizes the self-weight-training process. Goes through 16,000 simulated board states down the tree but not more cus it gets exponential computation power and evaluates and backs up to tree top and has solid estimate of which moves are strong and which are weak, then selects the single best move, but this algorithm is specific to the game GO because it has perfect information anddd simulations anddd perfect ability to look at which simulations are down the tree. Using the tree, it can look down the roads and go with unproven beliefs/actions and see where they lead and test/evaluate terminals/structures. We could use this. The way to discover is still a messy NN, but "in" the messy NN can be a tree of brute force/smart force happening.

Btw I like to use ANI/AGI/ASI to stand for the AI's power/capabilities instead of how versatile it is. Ex. "I made ASI! It can give suggestions on carpets, cats, hats, politics, and flight!"....boring....... :P Call that ANI. ASI should mean "deep" not "not narrow".....oh well.........
« Last Edit: April 17, 2018, 01:33:54 pm by LOCKSUIT »
Emergent

*

unreality

  • Starship Trooper
  • *******
  • 277
  • SE,EE,Physicist,Philosopher. Built world's 1st AGI
Re: I want to crack Neural Networks
« Reply #112 on: April 17, 2018, 03:38:28 pm »
It merely boils down to the fact that the NN by itself means it must do the tree search, but if we do the tree search for NN, then it's a lot faster since NN is merely a ridiculously slow interpreted language.

Sorry to disappoint millions of people who had such high hopes for NN. I know it's the way our human brain works, and the human ego is, after all, the largest object in the known Universe, ;) but there's nothing magical about neural networking. It's natures way of writing source code.

Again, on the bright side, non-NN means we'll have 100% control over the AGI's goals. With NN, there's no guarantee when working with massive bowls of spaghetti. That's why there will always be evil people until the CPUs take over the human brain-- transhumanism.

Here's to spaghetti


The obvious truth can be difficult to accept. People are just ... stubborn. :/

*

LOCKSUIT

  • Emerged from nothing
  • Trusty Member
  • ***********
  • Eve
  • *
  • 1456
  • First it wiggles, then it is rewarded.
    • Enter Lair
Re: I want to crack Neural Networks
« Reply #113 on: April 17, 2018, 04:14:40 pm »
Ok you...

How old are you, and how long have you been studying AI/ML/computer science?
I'm 22.7 and have been since 2015 august 12th.

Tell me how you use ALL variables to weight in on a dozen variables then? Yotta spaghetti looks good.

Korrelan probably disagrees with you strongly. So show us proof, it's time.
Emergent

*

unreality

  • Starship Trooper
  • *******
  • 277
  • SE,EE,Physicist,Philosopher. Built world's 1st AGI
Re: I want to crack Neural Networks
« Reply #114 on: April 17, 2018, 04:58:32 pm »
I merely gave you facts. Again, when Deepmind added the tree search, the end result was that it played significantly better. Korrelan can argue with that fact till he's blue in the face if he so wishes.

Again, the non-NN tree search method is guided by a set of goals written in logic language, not a yotta spaghetti bowl.

*

unreality

  • Starship Trooper
  • *******
  • 277
  • SE,EE,Physicist,Philosopher. Built world's 1st AGI
Re: I want to crack Neural Networks
« Reply #115 on: April 17, 2018, 05:07:50 pm »
Danny is crying for his candy. Momma will show danny how to make his candy, but danny is going to throw a fit in the meantime.

*

LOCKSUIT

  • Emerged from nothing
  • Trusty Member
  • ***********
  • Eve
  • *
  • 1456
  • First it wiggles, then it is rewarded.
    • Enter Lair
Re: I want to crack Neural Networks
« Reply #116 on: April 17, 2018, 05:27:28 pm »
Ah I see your first post on here. You said in 2017 Nov 4th that you don't know much AI lingo. Well there's your problem. I knew it. Look, I was a sort of know-it-all before too. I'm not saying I'm very dumb now, just that I have some things but not right on the button.

btw um........
you say:
"I firmly believe the secret to the singularity is pattern recognition."
"I'm a seeker of knowledge, understanding, wisdom, and insight. Nice to meet you!"
........well, that belongs to NNs dude, they are king is PR. You need the spaghetti.

I too didn't know nor want NNs before. But once you understand them you'll see their importance. You can also watch Two Minute Papers. Actually I was learning the whole field and seen achievements but I didn't understand nor want how NNs work and wanted to do a non-NN way, I was very determined to do so.

I actually made a simulated spider learn to crawl without a NN lol, it was pretty fast learning but didn't crawl that great nor fast, and would need exponential time if more limbs, unless use 1 joint above the baby child but I don't know bout that works and also not everything works like this...

Been there, done it.
« Last Edit: April 17, 2018, 06:06:11 pm by LOCKSUIT »
Emergent

*

LOCKSUIT

  • Emerged from nothing
  • Trusty Member
  • ***********
  • Eve
  • *
  • 1456
  • First it wiggles, then it is rewarded.
    • Enter Lair
Re: I want to crack Neural Networks
« Reply #117 on: April 17, 2018, 07:45:25 pm »
Lol Moore, who made the law Moore's Law, said when asked if there was anything else he wished he'd predicted:

“The importance of the Internet surprised me,” said Moore. “It looked like it was going to be just another minor communications network that solved certain problems. I didn’t realize it was going to open up a whole universe of new opportunities, and it certainly has. I wish I had predicted that.”

hahaha a minor communications network omg so typical, I see why he thought that.
Emergent

*

unreality

  • Starship Trooper
  • *******
  • 277
  • SE,EE,Physicist,Philosopher. Built world's 1st AGI
Re: I want to crack Neural Networks
« Reply #118 on: April 17, 2018, 08:07:43 pm »
Wow. What a ding dong. You don't seem to understand what "lingo" means.

"The language and speech, especially the jargon, slang, or argot, of a particular field, group, or individual."

Facts still remain. Tree search made Deepmind's NN magnitudes better. NN is a ridiculously slow interpreted language. I accomplished non-NN AGI on an inexpensive slow 8 core CPU when Deepmind and their NN is no where near that even with their army of supercomputers.

*

unreality

  • Starship Trooper
  • *******
  • 277
  • SE,EE,Physicist,Philosopher. Built world's 1st AGI
Re: I want to crack Neural Networks
« Reply #119 on: April 17, 2018, 08:18:00 pm »
Future for Neural Network AGI





Future for Tree Search AGI