Blade Runner 2049

[goDel]

I think you should read up on the newer alphago version:

https://www.inc.com/lisa-calhoun/google-artificial-intelligence-alpha-go-zero-just-pressed-reset-on-how-we-learn.html

 

Also, i get the impression that 'stateles function' might be personal jargon, or otherwise jargon which might be from a different field. In AI autonomous agents are 'stateful' almost by definition, i'd argue. Almost, because less complex agents might get away with a 'stateless function' I'm guessing. A self-driving car would probably need to learn as it goes. And continuously improves it's states. But I'm not sure whether I'm using your concepts as you would define them. Think I get the general idea, though.

[zlemflolia]

I think you should read up on the newer alphago version:

https://www.inc.com/lisa-calhoun/google-artificial-intelligence-alpha-go-zero-just-pressed-reset-on-how-we-learn.html

 

Also, i get the impression that 'stateles function' might be personal jargon, or otherwise jargon which might be from a different field. In AI autonomous agents are 'stateful' almost by definition, i'd argue. Almost, because less complex agents might get away with a 'stateless function' I'm guessing. A self-driving car would probably need to learn as it goes. And continuously improves it's states. But I'm not sure whether I'm using your concepts as you would define them. Think I get the general idea, though.

 

I'm probably using "stateful" in a non-rigorous way.  But for instance instance simple neural nets are stateless because there is no internal state to keep track of while executing them i.e. they can be executed while stored in const memory.  There is no state to keep track of and each scoring is deterministic. 

 

If it's trained ahead of time once and then an immutable instance is executed from then on, it's stateless

If it's trained ahead of time and also continues training itself during use, it's stateful

 

AlphaGo Zero in this case seems stateless by these definitions because it was trained ahead of time using first principles, rather than using actual game data (and presumably therefore also not using realtime in-game data apart).

 

Note the "state" of the board is not related to the state of the AI, it's simply an input and in this case it would be a stateless function of BoardA -> BoardB

 

A self driving car keeps track of lots of external state and maps that external state to transitions in internal state, but it would depend on the implementation whether the cars continue to learn forever on their own, or whether the cars send their observational data back to the car manufacturer, which is then processed into new ML models which are then sent back to the car and executed statelessly

Edited October 26, 2017 by Zeffolia

[zlemflolia]

EDIT:

Basically I'm using "state" here as the "personal memories" of the AI which influence future behavior.  Take two identical AIs and put them in different situations and leave them for a few hours.  Then bring them back together again.  Are they different from each other?  Then they are stateful.  But most "AI" systems we have now would not be different, because they are simple stateless models being executed on a discrete time quanta worth of input data, maybe a single voice command from a user (siri) or a single web request (Google search) which is then later processed into new state and updated after a larger discrete time interval

 

A self driving car keeps track of lots of external state and maps that external state to transitions in internal state, but it would depend on the implementation whether the cars continue to learn forever on their own, or whether the cars send their observational data back to the car manufacturer, which is then processed into new ML models which are then sent back to the car and executed statelessly

 

Basically I'm using "state" here as the "personal memories" of the AI which influence future behavior.  Take two identical AIs and put them in different situations and leave them for a few hours.  Then bring them back together again.  Are they different from each other?  Then they are stateful.  But most "AI" systems we have now would not be different, because they are simple stateless models being executed on a discrete time quanta worth of input data, maybe a single voice command from a user (siri) or a single web request (Google search) which is then later processed into new state and updated after a larger discrete time interval

[goDel]

The difference between "stateless" and "stateful" is in technical terms only a small difference. Do you train a static model, or is your model dynamic? A static model becomes dynamic of you simply keep on giving it feedback (or reinforcement). In terms of complexity there's not a big difference between stateless and stateful. 

 

A simple stateful example would be google translate. It works on a model which is continuously updated. If only by adding more text to the corpus, which it uses to model language.

 

alphazero go can be seen as a transition into what you'd call stateful. simply because it's capable to train itself without human intervention. you have to actively make it stateless by pulling a switch which says "stop learning". Until you pull that switch, it's stateful, in your terms, I'm guessing.

[Echolalia]

Saw this again last night in Shrewsbury. Fella behind me as soon as the credits roll "That was a bit shit that werent it?"

 

Errrrm no.

[goDel]

 

Basically I'm using "state" here as the "personal memories" of the AI which influence future behavior.  Take two identical AIs and put them in different situations and leave them for a few hours.  Then bring them back together again.  Are they different from each other?  Then they are stateful.  

 so a genetic algorithm would be stateful?

 

http://boxcar2d.com/

 

different results guaranteed. :)

[zlemflolia]

The difference between "stateless" and "stateful" is in technical terms only a small difference. Do you train a static model, or is your model dynamic? A static model becomes dynamic of you simply keep on giving it feedback (or reinforcement). In terms of complexity there's not a big difference between stateless and stateful. 

 

A simple stateful example would be google translate. It works on a model which is continuously updated. If only by adding more text to the corpus, which it uses to model language.

 

alphazero go can be seen as a transition into what you'd call stateful. simply because it's capable to train itself without human intervention. you have to actively make it stateless by pulling a switch which says "stop learning". Until you pull that switch, it's stateful, in your terms, I'm guessing.

 

Fair enough, I'm by no means an expert here.  My reading was that actual game data and therefore player data is not used by and therefore during actual gameplay no modifications to state occur either.  And it's a matter of update refresh time really, I highly doubt Google translate is literally trained in realtime where if I make a query, training differences are propagated immediately to the point where the servers have to keep track of a realtime state of incoming queries.  It would be batched and retrained, then new models would be propagated outwards through the network at varying time intervals.  So it would be stateless still in this case.  So this ML model refresh cycle could be analogous to the organic sentient being's flicker flash rate potentially

[zlemflolia]

 

 

Basically I'm using "state" here as the "personal memories" of the AI which influence future behavior.  Take two identical AIs and put them in different situations and leave them for a few hours.  Then bring them back together again.  Are they different from each other?  Then they are stateful.  

 so a genetic algorithm would be stateful?

 

http://boxcar2d.com/

 

different results guaranteed. :)

 

 

 

It would still be stateless by my reckoning since it keeps no memory of past generations, only the current one.  But that's state I guess, but not very useful state. 

 

I think my distinction doesn't work actually and needs more refining.  There are many types of state, I would say the current generation of a genetic programming algorithm is analogous to the current state of the board, not the current state of the AI.  There is nothing traditionally associated with AI in genetic programming apart from being a tool to search parameter spaces for useful outputs.  Genetic programming is not in itself useful for decisions, it's useful for creating tools that can be used in the future, essentially a training method.  The act of training the AI would be done offline ahead of actual usage and this is what genetic programming in the midst of occurring is, like this simulation

Edited October 26, 2017 by Zeffolia

[goDel]

that's an awkwrad logic. that's like saying that evolution keeps no memory of past generations. it's not a literal memory. but human memory is certainly not literal either. so the whole notion of memory is a bit off. 

 

also, an artificial neural network can be given a simple kind of memory by making a recurrent connection (technical at t1 the network also uses information from  the previous state t0. 

 

last, the distinction between continuous time and discrete time is really not that relevant. just like digitising an analogous audio signal. if you play it back, it's just the same. only discrete in a for the human ears impossible to hear way.

[Rubin Farr]

Saw this again last night in Shrewsbury. Fella behind me as soon as the credits roll "That was a bit shit that werent it?"

 

Errrrm no.

In my experience, polarizing reviews are a mark of greatness in art.

[Sally Vingoe]

just saw it today

i really liked it

maybe low expectations, I mean, imagine how bad it could have been  (imagine if ridley got his hands on it)

[Echolalia]

 

Saw this again last night in Shrewsbury. Fella behind me as soon as the credits roll "That was a bit shit that werent it?"

 

Errrrm no.

In my experience, polarizing reviews are a mark of greatness in art.

Inland Empire would suggest this to be the case...

[dumplings]

[olo]

dear god...

[Braintree]

Why do nerds need to have a pissing contest at every turn?

[auxien]

lol

[o00o]

[o00o]

 

[Echolalia]

 

Oh to be young again! The camera would certainly avoid my old self if I was there. 

[kebabkillaz]

ive seen the movie few days ago, cant get it out of my head. MASTERPIECE. if you think differently you are simply wrong and you should cease to exist

[Braintree]

ive seen the movie few days ago, cant get it out of my head. MASTERPIECE. if you think differently you are simply wrong and you should cease to exist

 

 

Read the rules.

[kebabkillaz]

i did, why?

[doublename]

Is it really pronounced “q-dough”?

[dumplings]

Is it really pronounced “q-dough”?

I thought the same. I always pronounced it kweedough, like Guido.

[sweepstakes]

Kway dough