Print

Please login or register.
1 Hour 1 Day 1 Week 1 Month Forever
Login with username, password and session length

[nenjin]

IIRC the Target algorithm looked at spending habits and matched those with others who tend to have babies soon after displaying similar spending habits. And so sent out coupons to people and they're like "WTF, how did Target know we're putting some bread in the oven?!!"

[Trekkin]

IIRC the Target algorithm looked at spending habits and matched those with others who tend to have babies soon after displaying similar spending habits. And so sent out coupons to people and they're like "WTF, how did Target know we're putting some bread in the oven?!!"

Yes, this. All machine learning does is find patterns, but that's not nearly as limiting as it sounds.

[smjjames]

So basically, predictable people are predictable and it fell into that uncanny valley of "WTF? Are you reading my mind?"

[Antioch]

All current machine learning algorithms are nothing more than just correlation-resolving. Finding the smallest subset of a data sample most correlated. Deep learning is just an alteration of a neural network structure still otherwise employing the exact same algorithms.
(...)

So is human intelligence.

While delightfully cynical, not really true at all.

So what is the difference according to you?

[MrRoboto75]

So basically, predictable people are predictable and it fell into that uncanny valley of "WTF? Are you reading my mind?"

Just plug the controller into the 2nd player slot.  No more mind reading!

[Antioch]

All current machine learning algorithms are nothing more than just correlation-resolving. Finding the smallest subset of a data sample most correlated. Deep learning is just an alteration of a neural network structure still otherwise employing the exact same algorithms.
(...)

So is human intelligence.

While delightfully cynical, not really true at all.

So what is the difference according to you?

I already mentioned it above, sort of. I'd say that the ability to build associations is just one key aspect of intelligence and not the entire thing. Key aspect, yes, but there's a lot more than just that. For example, if we take the usual example of language. Languages are learnt as an association between objects and the sounds used to describe them; generally these source from sensory information. The problem is when you get to the learning of words that aren't one that can be explained through simple sensory impact. Try finding an easy association-related definition for "cynicism".

There are other aspects as well, though that'd make this post considerably lengthier.

Neural nets can deal with abstract concepts like cynicism just fine, in fact almost the entire point of neural nets is abstraction.

Sensory input gets abstracted to concepts like "duck or chair" just like certain inputs are abstracted to things like beauty or value.

The deeper you go the more abstracted information becomes.

Concepts like cynicism can be formed from previous abstract concepts, like negativity or the nullification of positive value judgements and whatever other concept you would like to describe cynicism.

[SalmonGod]

Just plug the controller into the 2nd player slot.  No more mind reading!

[EnigmaticHat]

...have you ever actually used a neural net? I'm not entirely sure you understand exactly what a neural network does. It's not "thinking". It's not "abstracting" in any sense of the word. It's just taking a set of inputs and figuring out whether they yield a positive or negative output (true or false), then alters the weight of each path from node to node based on what was more or less relevant to the conclusion of true or false. It effectively takes from an input and isolates the most relevant details from each piece of data across an entire data sample to yield a proper output for each (when proper output is known), as to be able to figure out the output without knowing it beforehand.

I don't know where you really get abstraction from. At least, not in the sense that you're thinking.


Seriously, they're not magic.

This^

A neural net has an input and an output, that's all that it is.  Your AI for inventing random MTG cards isn't going to suddenly understand what a control deck is, much less start questioning the nature of love.

The rule with computers is that its always going to be many times harder to make a computer do something than it would be to do it yourself.  The computer may be able to do that thing faster, but its rarely going to do it better than its creator could.  A good example of this would be the mutual fund run by an AI (forget the name).  All it does is subscribe to every finance magazine under the sun and harvest them for keywords.  It might seem like it can out think humans, but its far dumber than a human even at the stock market.  What it does is read faster than a human ever could, its understanding of what its' reading is 100x worse than a human's would be.  But since no human could read all that text, it can compete from a position of greater knowledge.  The stock market (much like a casino) is also a place where emotions and creativity are liabilities instead of assets, hence computers having an advantage.

Think of it in terms of video games.  3D art is nothing more than a computer drawing 2D art at the speed of light.  But to get to that point, you need to be able to draw the 2D art.  Because its much harder to make a computer draw for you, than it is to draw in the first place.  If you couldn't draw the concept art, you can't make a 3D model and rig for your video game character.

So the issue becomes this.  In order to make an AI that can think better than a human, you need to be smarter than a human.  Not just a little smarter than a human, not just a genius.  You need to be like 10x as smart as a human.  We ourselves can't (yet?) make an AI that's as smart as an animal.  For some definitions of smart, we arguably haven't made an AI as smart as an ant.  All we've achieved are AIs that think *quickly*.  There is most definitely no hard evidence that AIs are ever going to be truly "thinking" instead of just calculating quickly.  That isn't to say that AIs won't come for our jobs.  I mean hell they already drive better than we do.  But they can do that because they're having google maps and a ton of visual information fed directly into their brains.  If I had that and could think at the speed of light, I'd drive well too.  So in terms of an AI running the government/world, how do you give it everything?  How do you feed a world's worth of contradictory and subjective political information into a computer?  Politics isn't like driving or even the stock market, its a much more social and creative field.  Not that we would think of it as creative but politics is very much right brain; there's no clear cut solution and you have to cut your own path rather than checking the right multiple choice box.  And the rules of politics constantly change as the situation does; there's no way that an AI to run the government 50 years ago would possibly suffice now, much less 200 or 2000 years ago (pretending that wouldn't be anachronistic).  So even if we somehow made an effective AI for governing we'd have to let someone modify it, re-introducing the fallible human element.  Of course machine learning algorithms could be a scary powerful *aide* to politicians, but you're not going to want to plug something like that into the nuclear silos.

[Reelya]

Neural networks are a whole other kettle of underdeveloped fish. Presently, the biggest neural network ever created BY a human has perhaps 100 neurons. The simplest neural system found in nature has 300 neurons and over 7000 synapses, and we can't even replicate it a little bit.

Uh what? Just googling the biggest NN ever created shows how far off that is.

https://spectrum.ieee.org/tech-talk/computing/software/biggest-neural-network-ever-pushes-ai-deep-learning

That one has 160 billion parameters. Each parameter is equivalent to one synapse.

also this:

http://www.telegraph.co.uk/technology/10567942/Supercomputer-models-one-second-of-human-brain-activity.html

It used the open-source Neural Simulation Technology (NEST) tool to replicate a network consisting of 1.73 billion nerve cells connected by 10.4 trillion synapses.

People have been saying 100 neurons is the limit for a long time, I heard that back in the 1990s. Forgetting Moore's Law there ... and it was probably innacurate even then. You can have as many neurons modeled as you have RAM. It's not proper measure of processing power.

[Trekkin]

I'd argue they could also be a scarily powerful aide to governance, though. Computers have two additional advantages over humans: they can't accidentally forget things and they are fast enough that we can regularly get useful output from them without overtraining to the point of never seeing anything unexpected turn up (or, at least, the degree of enrichment vs. diversity in the solution set is tunable.)

Admittedly, my networks fold proteins rather than drafting policy, but I can see ways of making a network to produce first approximations of the numerical values that so often produce unintended consequences, like all the step functions in our welfare programs that discourage people from seeking raises by taking more benefits away than the additional income could replace. You would definitely want people to screen the output, as with anything a computer generates, but all the obstacles I can see between us and having a computer narrow down tons of possible values to something human-scannable are more questions of optimization than possibility -- or are external to the computation itself, like with Ispil's hypothetical concerted efforts to feed garbage data into the machine.

So yes, weird's techno-dictator is neither possible nor a good idea, but I could see some kind of cyborg CBO being a potentially useful guard against legislation not doing what it's designed to do simply because it's complicated and fiddly and we know more about what we want than about how to get there.

[TheDarkStar]

So rather than "humans need not apply" we get "non-augmented humans need not apply".

[Reelya]

AI can definitely write bills. It would probably do better than humans at that. Lawyerbots are already working for law firms, replacing research clerks. I hear journalismbots are already writing a significant amount of the text you read in news too. It's not too hard to see that logical inference systems could avoid common pitfalls of badly written bills.

https://www.wired.com/2017/02/robots-wrote-this-story/

[Trekkin]

So rather than "humans need not apply" we get "non-augmented humans need not apply".

I was being metaphorical. I meant "cyborg" at the organizational level; machines create the alternatives and prune the obviously ridiculous ones while humans review the remainder. No part of that process requires any more sophisticated interface between the two than a monitor and a keyboard.

[Reelya]

The whole thing with Generative adversarial networks (developed as recently in 2014) is a game-changer in what we can do as well. e.g. the pruning algorithm you mention isn't some inert thing that never changes, it evolves in response to human decisions, so it's always getting better, and you cycle back and forth between the "policy creator" and "policy rejector" networks possibly thousands of times before a bill ever needs to be evaluated by a human. The training goal is that your "policy rejector" network is trying to mimic what the human would decide, and the "policy creator" network tries to generate bills that would fool the rejector.

[Antioch]

...have you ever actually used a neural net? I'm not entirely sure you understand exactly what a neural network does. It's not "thinking". It's not "abstracting" in any sense of the word. It's just taking a set of inputs and figuring out whether they yield a positive or negative output (true or false), then alters the weight of each path from node to node based on what was more or less relevant to the conclusion of true or false. It effectively takes from an input and isolates the most relevant details from each piece of data across an entire data sample to yield a proper output for each (when proper output is known), as to be able to figure out the output without knowing it beforehand.

I don't know where you really get abstraction from. At least, not in the sense that you're thinking.


Seriously, they're not magic.

Isolating relevant details in this way is precisely what abstraction is.

Neurons are also just taking a set of inputs that are figure out whether or not they yield a a positive or negative output with a weight associated to each path.
The only fundamental difference is scale. What other "aspects" does an animal brain have, not found in a neural net?