Machine learning is hard

because machine learning (neural networks), i.e. a real software technology that's deployed by google, apple, etc. has absolutely nothing to do with quantum computing (which is basically a meme among academics who get government grants because "muh cryptography".)

(unironically cryptography is important, e.g. look at WW2 and how codebreakers won the war... but in quantum computing, they've basically not accomplished anything but still use the buzzwords that attract the big buxx that politicos like for delegating defense spending)

Universal quantum computers = greater speed to train neural nets which will take a tremendous data for applications like translation. That's exactly what it is. Stop pretending like what you think you know is deep or complicated. It's embarrassing.

Attached: 1528115140443.jpg (707x504, 68K)

thanks nigga u a real wan you feel me that's some real ass shit bruv I fucking love apple and shit u smart nigga

usa mssart ass nigga u feel me, love you bro

youtu.be/SO1ntFh5VuE

Attached: download (21).jpg (275x183, 6K)

oh shit, sick beats.

anyhow i stand by my claim that:
> neural networks = real technology people make $$ with
> quantum computers = hypothetical technology academics get money with by shilling buzzwords about cryptography to the DoD

btw this song is sick

Interesting

It's 6.47 in the morning so I'll keep it brief.
In short:
>Literally can't be possible
False, neural networks have been shown to be universal function approximators and the progress in transfer learning (models that apply skills learned for a task to another task) is going pretty fast.

>We have quantum level computing
Assuming he was talking about quantum computing, then a) there are maybe 10 useful quantum algorithms in the world and all the rest are just info theorists masturbating, b) quantum machine learning (e.g. quantum perceptron) has been proposed but it has literally zero advantages when considering that a quantum bit costs about 10k$ and c) not once in the last 10 years someone has said "oh cool algorithm, if only we had a QC available we could finally run it".
Also this post is bullshit because universal quantum computing with crazy speedups yayyy! is not a real thing. You get exponential speedups if you can devise an algorithm that exploits quantum interference, that's it.

>A quadrillion books
First of all, we have basically all of the books ever written neatly digitized and paired for translation, you just need Google's manpower to exploit this data. Second, building a human MT model is not really about data, as it is about smarter/more expressive algorithms (better to use an attention network than a 1000000 layer network trained on a gazillion ).

Attached: 1520876943249.jpg (720x736, 94K)

>He thinks quantum computers will have the parameter space needed to train neural networks for Natural language processing or computer vision.
Also, google translate works pretty well so I'd say the computer translation problem is near solved.
Generally its not the computing power thats holding us back, its our algorithms. We currently have a brute force method (Neural Networks) that work pretty well but IMHO arent the right method to solving difficult machine learning problems problems. First we need know how the 'universal function approximator' in our brain works, if such a thing exists.

>Wanting a machine to learn enough to kill the human race
Are you retarded? How can you be smart enough to program AI stuff but be too dumb to see the future consequences of this?

Just put neural networks on the block chain and power it using the ethos of Web 3.0 on the cloud it’s not hard

>because machine learning (neural networks), i.e. a real software technology that's deployed by google, apple, etc
>neural networks = real technology people make $$ with
You've never touched a deep learning code in your life. I'm not even gonna
>hurr it doesnt use 0 and 1 its a meme
because even that would be much beyond your level. Tell me, why are you pretending like you know shit? I'm 90% confident you don't even know what backpropagation is.
>neural networks have been shown to be universal function approximators
Ask me how I know you don't have any idea what the fuck you're talking about.
>hurr just feed the computer books and let them find the pattern
ITS NOT EVER GONNA WORK YOU FUCKING BRAINLETS
The computer does not UNDERSTAND what it's doing. You need more data than you can possibly imagine to let the AI mimic results of human level cognition (which is required for tasks like translation), because it's basically the same thing as brute force cracking though in the end it might make mistakes. God you brainlets make me sick.
>translation problem is near solved
Oh no no no. The AI cannot possibly tell things like sarcasm without really understanding what it's talking about. Simply put, one neural net isn't enough.
>hurr

Attached: 1517287176345.png (213x237, 8K)

This is exactly what I mean. The cavalier attitude where you act like a teenage that doesn't consider the consequences of their actions.

Attached: machine_learning.png (371x439, 28K)

If you don't understand how AI works why don't you just go learn the basics instead of "hurr its going to kill humanity"? Because you have no idea what you are talking about, and that is a fact. When you actually learned how neural net works you'd feel embarrassed to have even asked a question like that.

Attached: 1531537768058.jpg (1280x720, 129K)

>machine learning equals artificial intelligence
>singularity in 2045

Attached: i seriously hope.png (295x295, 2K)

Once TPU 3.0 based AI supercomputers hit the speedup would be tremendous. If the current progress continues we might indeed see computers besting humans at most tasks by 2050.

so my nigga uz is sayin dem niggas in da commputer is like functining out all them shits n shit. That's crazy. Then yo niggaz use dem numbahsz to make tdem numberz make the computer talk asian.

Damn

Are you objecting to the fact that NNs are universal approximators or that a universal approximator is needed to mimic human intelligence? Because in either case you're wrong.
Also, why do you think that self cognition is necessary to translate a language to another? It's barely a matter of grammatical rules and knowledge of non-literal expressions (e.g. fucking around != Actually going around to fuck).

>human level AI within 30 years
"""experts""" said the same thing in the 60s

Attached: average robot 2018.webm (720x404, 1.81M)

Where?

>The computer does not UNDERSTAND what it's doing.

So? Niggers do this too, and they approximate real humans pretty well, fooling the whole mainstream establishment.

NN is just a big correlator, just like human brain.

The optimism

The first generation of AI researchers made these predictions about their work:

1958, H. A. Simon and Allen Newell: "within ten years a digital computer will be the world's chess champion" and "within ten years a digital computer will discover and prove an important new mathematical theorem."[59]
1965, H. A. Simon: "machines will be capable, within twenty years, of doing any work a man can do."[60]
1967, Marvin Minsky: "Within a generation ... the problem of creating 'artificial intelligence' will substantially be solved."[61]
1970, Marvin Minsky (in Life Magazine): "In from three to eight years we will have a machine with the general intelligence of an average human being."[62]

>It's barely a matter of grammatical rules and knowledge of non-literal expressions (e.g. fucking around != Actually going around to fuck)
How does the AI know the context of this, that it means they aren't literally going to fuck?
You need one billion books where people actually fucked in one certain context, and another billion books where it's only figurative. That way it may appear the AI "knows" something about the world, whereas it acutally doesnt.
Imagine you only have 2 brain cells, one consists a giant neural net like this, and the other one is memory. That's your "AI".

Attached: 1527068269416.png (548x666, 361K)

this guy is definitely some sort of larper. ignore him.

of course, minsky was trying to get the $$ from the DoD.

how do you think he made his living?

but can you program a machine learning algorithm to get laid?

Attached: 1527944621138.jpg (1125x1341, 800K)

neural networks are good at finding approximate solutions to nonlinear equations, which normal techniques are not that great at

that doesn't mean ANYTHING about AI.

the fact that even in cognitive science all the top dudes are like "WTF IS CONSCIOUSNESS? IT DOESNT EVEN HAVE AN OS" makes it clear that even the best NNs are decades, even centuries, away from AI

run a chatbot on tinder

We could build him.. the most alpha male. With the most chad bio, computer generated chad face, and game of an absolute monster. using location emulation and thousands, no, millions, of tinder accounts.

How many generations would we need to make him, the Chad, what would be the maximum "willing to have sex" rate. We then, genetically engineer a sex god. Olympus him self.

GET TO WORK LADS.

Machine learning is like making a pasta strainer

You take a shit ton of wires (aka data) and you arrange them in a very nice way (adjusting it's weights) and in the end you walk away with something that can remove pasta (but only pasta)

If we had as much data as your implying we'd already have a distribution to sample from. The point of machine learning is to infer patterns when you dont have a distribution to sample from.

>m-muh singularity in 20 years

Attached: singularity.gif (504x716, 68K)

And that's what makes it impractical for many tasks because reality is abstract and you cant have enough patterns for everything without that much data. The AI can only guess within a very limited scope, nothing more. For example, the AI cant write a book and be expected to win nobel prize because all it does is putting words that are used frequently in other books together without considering their meaning.

Attached: 1527408225245.jpg (412x635, 77K)

> plagiarism is like hard work, but easier

Yes. We need to replicate antennae for the idea world.

Epic post m8.