/sci/ - Science & Math

>>11279892
>Machine Learning
>"""""""AI"""""""

>>11279892
>Tesla AI is already hitting the top because it still cannot distinguish between parked cars and moving cars during normal driving conditions and it cannot possibly work on its own when there is fog, rain or snow.
WTF are you talking about

>>11279892
Chinese credit score system

>>11279902
Teslas don't know whether to choice between moving and stationary cars when the road is narrow and there are parked cars on one side.

Tesla AI cannot navigate through fog, rainy or snowy days on their own, there is a different system to handle those situations, I think it's sonar/radar or some shit.

>>11279913
Sounds like you are making up random nonsense.

>>11279947
Please stop being such an idiot fanboy.

>>11279892
>AI
That is a popsci/sci-fi term and no such thing exists yet.

>>11279892
nobody expected machine learning to become a conscious program to begin with, there will be another AI winter after people notice that ML will get us no where, hopefully it will be brief and we will have full brain simulations at the end of it, until we can do a brain sim there is no hope

>>11280082
>until we can do brain sims there is no hope
why? I'd heard that at this point neurocomputational mapping is more for extracting useful possible architecture and looking for redundancies to avoid when mimicking real nervous sytems. Otherwise its ostensibly inefficient to try to approximate natural selection as it does not optimize for raw veracity or computing power.

>>11279892
It's clear from your writing you don't know shit. Maybe learn some things about "AI" before posting here?

>>11280255
I think I over did it a little when using the term AI so losely, but in the end it's just a bunch of retarded linear algebra that is pretending to be smart when it is ultimately going nowhere anyways.

>>11280270
No, it's not that, you are for sure just a moron.

>>11279913
> i think its sonar/radar or some shit

Why do brainlets think they wont be called out? Half of this board’s threads are made by people with absolutely no knowledge to any of the topics they are conversing

>>11280281
You are either a Tesla faggot, a low iq instagram whore or a combinations of those. It doesn't matter, you are still racing to a red light

>>11280293
Teslas CANNOT POSSIBLY navigate using neural networks when there is fog, rain or snow BY DESIGN, that's a hard fact you stupid moron. It uses a different feedback subsystem, which I don't care to look for.

>>11280293
>>11280255
>>11279947
There you go, straight from the official Tesla website you room temp iq morons.

There is obviously a ceiling as to how much the models can be optimised and how much GPU power can be thrown at problems. If we pretend that the brain for example is a computer then it probably contains countless layers of nonlinear operations, which are probably out of reach of any known transistor-based computation. I haven't found any good reading as to where the actual limit lies.

>>11280346
Some researcher shed some light on the problem not a long time ago, he basically said old machine learning algorithms are on pair with some of the newest ones when there is enough raw power to crunch numbers. Basically machine learning has become a matter of brute force at this point and most of the latest developments aka algorithms are nothing but an absolute waste of time and money in the grand scheme of things.

http://www.incompleteideas.net/IncIdeas/BitterLesson.html

>>11280306
>>11280316
Most people in this thread are just laughing at you, and will continue to do so until you give up and this 404s, but let me help you out.

>Teslas CANNOT POSSIBLY navigate using neural networks when there is fog, rain or snow BY DESIGN
You seem to not understand what a neural network is. A neural network doesn't require images. A camera is a sensor, just like a radar. Radar data can be fed through a NN just like images. Also, if a human can see, a camera can see, so saying weather makes using cameras impossible is no different than saying it's impossible for a human to drive in rain or fog. Human's don't have radars equipped, not sure if you knew that.

>>11279892
I know a research scientist at Deepmind. I'm going to be as kind as I possibly can - if he's representative of this industry then I can see why AI isn't going anywhere.

>>11279980
>>11280300
Also, the "anyone who disagrees with me is a shill/onions wojack/etc." isn't effective when your own posts prove you to be incompetent, so if that's your strategy on this board, try to be more vague when posting instead of putting your lack of knowledge on display.

>>11280406
What stupid shit has he done?

>>11280400
Radar is not computer vision you absolute low iq tranny, in fact, that's not even the selling point of Teslas lmao. On a technical level, Teslas are marketed as computer vision solutions to tackle the self driving problem. That's it.

Holy shit, even Elon Musk says over and over again that Teslas are all about COMPUTER VISION. That's why Teslas shit the bed when they are navigating through low visibility situations, and that's why they need radar data to assist the main driving algorithm.

Keep trying though.

>>11280316
I wasn't taking about your Tesla ramblings when saying you don't know shit, but thanks for providing even more laughing material.
This Anon gets it
>>11280400

>>11280424
>radar is not computer vision
Holy shit, you really aren't trolling, lmao.

>>11280438
Let me dumb this down for you, ok?

Tesla "AI" computer vision experiment is the only thing in the entire world that represents a massive real use case for machine learning that could be life changing for all humans. If FSB happens, then machine learning wins, massive leap, tesla stock bullrun, blah blah blah. Easy right?

Now, said company is at the literal cutting edge of this meme technology. However, even after 6 years and probably Exabytes of information, Teslas cannot navigate through fog, rain or snow without helping wheels aka radar, a technology which is, according to Elon Musk, CEO of Tesla, retarded. Also, they cannot distinguish between parked cars and moving cars when the road gets narrow.

Tesla claims the only reason we don't have FSD yet is government regulations.

Now let me ask you this: why would a computer algorithm that has been feed with thousands of years of driving experience achieve full self driving, ever? And most of all, what does this say about Machine Learning in general?

Do you wanna make it in the world of machine learning? Learn how to make instagram tranny filters, because in less than a year this entire meme industry will crash and burn like there is no tomorrow.

You are welcome.

>>11279892
On a similar note, is OP still retarded as fuck? pic extremely related.

>>11280470
You're not wrong that AI is still bullshit, but you're 100% wrong that radar can't be used for machine vision. The algorithm doesn't care if the input is from a camera, lidar, radar, or a midget installed in the frunk who calls out what he sees with an obviously fake Italian accent. All that input can be used as part of a machine vision system.

>>11279913
>>11279980
>>11280424
>>11280470
Maybe you should go back to /pol/, because you can't seem to link any concepts together on your own that haven't already been done so step by step, then fixed in your head permanently and immutably forever as some sort of rule of thumb, by someone you choose to trust based on your emotions rather than any sort of objective correctness. That's not science. That's what science and reality at large is reduced to for imbeciles who are fundamentally incapable of engaging in inquiry and experimentation themselves. You know the bell curve? You're on the lower end of it. Enjoy.

>>11280541
Ok, then no matter the input type, machine learning still cannot solve for self driving. It probably doesn't even have a guaranteed solution and that's exactly my problem with this current trend, people who are into deep learning have started to develop more faith than hard cold solutions to tackle the problem. I'm pretty sure tesla engineers are in some deep shit because those hardware processors they are developing were basically a bet on neural nets being capable of everything by simple feeding them with large amounts of raw data.

>>11280601
lmao didn't even read your rant
clamped at birth that's for sure
dilate

>>11279892
suck nuts and eat cum
https://www.youtube.com/watch?v=DfznnKUwywQ

>>11280688
>machine learning still cannot solve for self driving
You can't "solve" for driving. There are way too many variables. Our brain hasn't solved for driving either. We too just approximate solutions to things we encounter in life, but we have the benefit of being born with billions of years of highly optimized training data.

>>11281141
>but we have the benefit of being born with billions of years of highly optimized training data.

we never did anything like driving cars in our evoloutionairy history, yet we can learn to drive cars after less than an hour of training

meanwhile self driving cars accumulate millions of hours of training, consume terawatts of computing power in the process and can't even approach fucking hairless monkeys

>>11281149
This thread is still full of laughing material, kek.
So having a huge backstock of 3D-vision and spatial perception training data isn't helping when learning to drive? Having witnessed drivers for 16 years? Games? Having cars optimized for human anatomy?

>>11281149
>we never did anything like driving cars in our evoloutionairy history, yet we can learn to drive cars after less than an hour of training
I swear some dumb freshman always makes this argument in these threads
it gets pretty tiring

>>11279892
self driving cars will be made irrelevant very quickly in the face of genetically modified horses.

>>11281149
driving is literally
>look and be perceptive of your surroundings
>then press two pedals and rotate a wheel accordingly
you have been learning the first part since your birth
the second part is simple and you can learn it in an hour

>>11279892
Generally machine learning is starting to hit the wall and diminishing returns have started. For the effort put in there’s less returns now.

Example:
https://www.wired.com/story/facebooks-ai-says-field-hit-wall/

>>11281239
Jesus Christ I know for sure you have never driven a real car in your entire fucking life. At a brain level, driving a car is an incredibly taxing task where conditions are constantly changing over time, non stop. Now get out and go back.

>>11281239
DNA powered computers have will to live so that makes things like driving easy, electric retarded computers are so fucking dumb that you can just turn it off and it'll just let you and it'll stay off even though it could turn itself back on if it wanted to, its got no will to live. thats why it has so much trouble understanding why driving into a brick wall at 120mph is bad.

>>11281281
>driving a car is an incredibly taxing task
lol no. You must be one of those fags who are scared to drive to the corner store. Track driving or riding while trying to beat your best is taxing. Landing a small plane in turbulent conditions is taxing. Driving on the highway is relaxing and driving in a city is frustrating just because people are retarded. As far as mentally engaging things go, I'd put driving somewhere around 2/10.

The. new cars will be more lidar and radar focused than CV.

>>11281297
Driving automatic cars isn't real driving, even retarded kids like you can do that. So you are telling me you have landed small planes now? LMAO ok lil Timmy, I think now it's time for you to leave your room and turn off your computer, your mommy probably left some food on the microwave, you massive larping faggot.

>>11281295
>DNA powered computers have will to live so that makes things like driving easy
Never before have I hoped this hard that someone was just pretending to be retarded. I don't want to believe this is an actual belief someone has.

>>11281328
>Driving cars isn't real driving,

>>11281365
>Driving isn't real driving

Please stop being retarded, zoomer

>>11281328
>Driving automatic cars isn't real driving
If you use a motorized starter, fuel injection, power steering or hydraulic brakes, it's not real driving.
Weak troll.
>So you are telling me you have landed small planes now?
Planes and gliders. Not everyone on 4chan is a pathetic basementfag. Don't project on me.

>>11281233
lol

I don't understand, are the morons here actually arguing that we aren't in for several decades of a AI/machine learning/computer intelligence winter?
Are you retarded?

>>11281530
we got amateurs at 12 o clock, check your safety!

>>11281563
The people here are probably so optimistically dumb that they think "w-we aren't hitting the end of moore's law... there are other paradigm coming!" lmfao stupid fucking dumbfucks

>>11281530
No, we're arguing that despite being 10% correct (probably by parroting shit a competent person said), OP is still a complete fucking idiot. I wouldn't be surprised if his posts were actually produced by an AI.

>>11279906
bost moar blease

>>11280406
Dated one of the few females at Deepmind's Edmonton campus. Interesting work, but buggy as shit

Granted I'm a family doctor so I know fuck all about AI

>>11281566
We hit Moore's law like a decade ago but we have Raja's law now

>>11281652
Never heard of that. Sounds pretty cool!

>>11281652
>Raja's law
So this is the power of Intel's state of the art all Indian marketing team. Amazing.

>>11279892
Nope. They're not even to the second floor, there's atleast few more floors to hit up.

>>11279913
Low IQ post. Whats out in public perception is not what's going on in underneath. Its not "they don't know about moving/stationary" Any NN can do that, not only that, any 10 minute NN can detect speed/distance/etc from moving images. The problem is how to manage those parameters into smooth driving experience. That's the hard part. You need bit of a holistic/contextual understanding for that.

>Tesla AI cannot navigate through fog, rainy or snowy days on their own
Its not an AI problem, its a visual spectrum problem. If light can't get to your camera/lidar/eye, then its a problem for any visual based system. It doesn't mean your brain is stupid.

>>11280406
Deepmind and openai are known to hype up their results without delivering and often receive negative criticism for this from other researchers. Still it is usually pretty hard to get into these labs and I doubt you can get there without being really smart.

Also could all the retards in this thread stop commenting if they know nothing about anything and just repeat popular researcher X's opinion from which they read some interview.

Most of the predictions made about AI including machine learning are completely worthless. Even if they are made by the most reputable experts, they are still just opinions. These fields move so fast and dynamically (or not at all) that no one can really accurately predict what will happen next.

Of course AI and especially machine learning are extremely overhyped at the current moment, but that doesn't mean no progress was made in recent years (a lot of it is just due to an increase in computational power though). And to assume that AI is stupid just because it's based on basic math is naive. For all we know the brain could do the same thing on another scale using other functions and optimization methods we haven't found yet.

>>11281652
>discrete AI as computing device
What did he mean by that?

>>11281656
that’s what happens if people actually study to get out of a shithole

>>11281757
Did you not realize I was making fun of that obvious wankery of a "law"? That shit is a marketing slogan, not a real phenomena.

>>11281630

>>11281233
Fucking based

>>11281677
I agree. I'm betting that if we see advancement in the area of computing and artificial intelligence it's going to come from the theoretical/mathematical side of things. I get the feeling that most researchers in machine learning these days don't get this.

>>11281652
>From CPUs and graphics cards to made up buzzwords that includes the word ai

Holy shit the absolute state of Intel india

>>11281791
Did you mean neuroscience, and mathematical models of the brain?

>>11281770
So when did we plateau?

>>11279892

AI is basically a fancy word for a general set of algorithms that are good at learning how to do incredibly specific tasks so that the developer doesn't have to write a massive, complex new program for that single specific task.

AI is good for picking out incredibly specific patterns at a level of detail that most humans fail to notice. For example, astronomers using AI to find something they missed in a dataset by teaching an AI what to look for. This can work against AI, since it analyzes data in such detail, slight variations that do not match the pattern it is looking for will cause it to fail to identify the pattern at all. The solution to this is to train with less data, and unmarked data, but that results in reduced confidences on patterns you want it to recognize. So now you need a human to intervene and examine everything it identifies with confidence greater than some amount, OR you just trust statistics and accept that 1% of the time your car is going to run over a pedestrian.

So AI is really good for any non-dynamic circumstance. Anything specific that you train it to do, it can do, and its certainly much faster to slightly modify an existing AI algorithm and teach it to do something, than it is to write a whole new algorithm specifically for that thing.

So once the Tesla AI has been trained repeatedly for ALL possible circumstances it will encounter on the road, then we will have self driving cars. However, the second it encounters anything it hasn't seen before, or even anything it has seen before but with a slight variation, it will fail to respond correctly.

I'm sure some mathematician could tell you the exact amount of training data it will take to achieve this, and I'm sure Tesla calculated that when they did feasibility studies on this. The fact that they are pursuing self-driving cars should tell you that it is possible withing a relatively short time horizon (<50 years) given the rate at which they expect new data to come in.

>>11279892
>>Where are those AI generated designs for cars and planes some TED talk retard was talking about ten years ago?
that was mostly topology optimization being way over-hyped. Topology optimization is used today, pic related was topology optimized. Topology optimization's getting pretty good and might take a big bite out of structural engineering jobs within 10 years.
>>11280985
it's crap. All they do is take a couple predefined linkages, generate a big database of curves for different, search the database for the closest curve, and optimize the parameters of the linkage. It only works for a couple predefined linkages and can't invent entirely new linkages. It only takes into account position and not forces so it's only good for animation. Go fuck yourself.

they also can't tell if a parked car is parallel or perpendicular to you when driving past them
tesla is finished

>>11281925
Plateau on what?
>Hardware: 2013
>Software: 2003
>Computer science: 1990
>Information theory: 1980
The last two might seem like approximates, but they're exact years.

>>11281932
>1% of the time your car is going to run over a pedestrian.

this isn't an especially hard problem to solve. your roomba probably knows how to avoid colliding with objects. i'm sure self-driving cars will avoid colliding with pedestrians at least as well as any human driver.

navigating complex traffic situations and planning effective strategies through congested traffic is likely a bigger challenge. stopping is easy, keeping a car moving is probably harder.

>>11282018
cont.

it'll be interesting to see how they deal with aggressive human drivers, and how aggressive it will be. legally, it has to yield right of way in certain situations. in reality, humans sometimes have to be a bit more assertive to get where they're going in a reasonable amount of time. what if 2/3 lanes are blocked and it has to merge with bumper-to-bumper traffic? will it inch into the correct lane like a human driver might, or will it perpetually yield as required by law?

>>11282030
cont.

and human drivers are very good at interpreting each others cues and signals in traffic. i think this is where they'll have the most problems.

>"shark"
>99.9% confidence

has actual intelligence hit a hard ceiling?

What if, and this may be a stretch but hear me out, the fundamental theorem of computer science is WRONG and the actual chemical and architecture of the machine is important for how a software can run on it?
What if the general intelligence of the human brain is literally simply not physically possible to run on the style of transistors that we design, there does NOT exist an abstraction, and it literally requires a biological cell based neuron in order for true general intelligence to exist on the information substrate?
Why is this not considered a valid possibility? I'm serious, I dont know and I"m really wondering. This is NOT SAYING that there is something "magical" about biological life, it would simply be a law of physics and computation, so there is no magic, you can't use that to counter this proposition.
I just want to know why everyone things you can just trade out any model of computation for another. The Church-Turing thesis has nothing to do with whether or not general intelligence has certain physical requirements so I don't understand why people act like a machine and a human brain are necessarily equivalent in terms of information processing.

>>11282084
Agreed
I believe analog computing is more suited to AI, but analog computers aren't as advanced as our digital ones. It was surpassed by digital computers and we never really looked back.

>>11281652
Why is GPU introduced so late? I also don't really understand why it's being pitted against CPUs and "Discrete AI". Actually, what even IS discrete AI? That makes it sound like there's dedicated AI cards or something.

>>11282084
>Why is this not considered a valid possibility?
Because it's been proven that Turing completeness implies the ability to solve all computable problems, and physics is computable. Some things are very slow on a classic computer (factoring a massive composite of two primes, for example), but we can always add a co-processor to handle those bits more quickly.

>>11282139
It's also been shown through Shannon-Hartley that analog systems are bounded by their noise floor. Brains are efficient because they're massively parallel and they're running some really well optimized algorithms, not because they have some analog component to them.

>>11282170
For now, AI processors are just simplified GPUs.

>>11282284
Is the perceptron a good model for the neuron though? In the perceptron, all the outputs have the same signal but is that true for a real neuron? Or is it possible that the axons (that might be in the hundreds of thousands) all have different voltages? And in that case the perceptron is just the dumbed down version of the neuron?

>>11282325
No, but it hardly matters. Neurons are not the only solution to an efficient learning machine. They're simply a solution that life fell into through evolution. There is no need to emulate neurons to make an actually decent AI.

>>11282334
>There is no need to emulate neurons to make an actually decent AI.
What are the other options?

>>11282325
Neurons are the combined function of RAM and a processor in a single unit. Every neuron can both store and process information. The voltages are not imporant quantitatively. Myelination and action potential propagation are better things to learn about.

>>11282084
Dude, there is a whole thesis about this by Roger Penrose, a physicist, based on the fact that we don't assume our computers are conscious, because they are simply equivalent to sticks and rocks bouncing producing an algorithm. His first book on this is The Emperor's New Mind, where he starts to grasp the problem, and the second one is Shadows of the Mind

>>11282341
Do you expect me to explain the entire ML field in a 4chan post?

>>11282369
Your brain is sticks and rocks if you buy that pseud trash.

>>11282362
That's not what I meant. A single perceptron has one output that connect to an arbitrary number of other perceptrons as theri inputs. But if real neurons can have a combination of outputs and connect to a thousands of other neurons then that makes the entire neural network a lot more dense and perhaps more efficient.

>>11282372
Sorry. But are you saying emulating neurons (e.g. neural network) is a dead end?

>>11282387
What?

>>11282391
The addition to your statement is that combination of outputs then becomes a storage point in itself which can then be connected as a single unit, or as an individual outside of it's subunit. That's the point of bringing storage + processing into the discussion.

>>11281652
>Raja's law

kek

>>11281652
>Pajeet's law
OH NO NO NO NO

>>11282084
Some high profile HFT trading firms use some sort of weird architecture where the computers don't operate on clock cycles but rather on some sort of continuous stream of data that is unbounded. They need that to operate at sub millisecond resolution when buying and selling their shit on the stock market. I think next level machine learning should operate under said conditions, because current solutions are probably already asymptotic

>>11282284
>Because it's been proven that Turing completeness implies the ability to solve all computable problems, and physics is computable.
How do we know this seeing as we do not have a TOE?

>>11282387
How is penrose pseud trash? he's smarter than every person in the field of AI

>>11282284
>Because it's been proven that Turing completeness implies the ability to solve all computable problems
low iq

>physics is computable
nevermind, you're retarded

>>11282543
How is he retarded?

>>11282547
first because turing completeness implies the ability to solve all problems computable by a turing machine. that's it. there's no proof concerning the question of whether there is a computation model more powerful than the turing machine or not.

second because we do not know every physical law of the universe, not even remotely. the ones we assume to exist just don't add up to create and maintain our universe.
so saying anything like that about physics does imply an iq below the level of retardation.

>>11282073

>>11281932
>the second it encounters anything it hasn't seen before, or even anything it has seen before but with a slight variation, it will fail to respond correctly.
this is the one advantage of neural networks over hard coded solutions besides too many variables. They are able to interpolate working solutions on new problems quickly.

obviously the future is in synthetic biology - that is, replacing the body's systems with machine equivalents. eventually you will just be a brain in the jar.

however, that's obviously expensive to keep you alive, so you need to have some way to earn money. well, an easy way to do that would be to run your support platform for someone else's benefit. you do something for them, and they provide you with nutrients and maintenance to keep you alive.

you will be kept alive as the brain to drive someone's car

>>11281369
>ANL

>>11282601
Holy fucking shrek I’m here for like 10 minutes and I get fucking goatse’d.
This board is brilliant I love you guys a lot

>>11282935
Dude is that some dark futuristic movie script or what

based and kinda unclamped

This threads shows just how far we’ve fallen from the hopes and dreams of the 80s. Compare Back To The Future with what we have today, and you’ll understand the despair of the millennial.

>>11281694
Anyone?

The real advancements aren't publicly known. AI is the new nuke. It's stupid and a huge national security risk to publicize that you have strong AI

>>11282481
If it's digital it has maximal resolution, it's never continuous. Show me proof of that statement, otherwise I'm calling bullshit.

>>11280028
Artificial intelligence is official term used by researchers and there already exists multiple AIs for many different domains.

>b-but AI means HAL9000 and Glados
Stop getting your information from movies/games and pick up a book from library.

>>11284421
>Artificial intelligence is official term used by researchers
Because "Progressive algorithm" wouldn't bring in as much investor money, so now we're stuck with using the term AI because it's convention

>>11279892
What if instead of tensors, they used sheafs...

>>11284312
I'm not saying it isn't digital, all I'm saying it doesn't have a clock

https://www.youtube.com/watch?v=OgS52RyiO_M

39:00

>>11285058
Just watched what you were talking about. He's just talking about systems reacting to sequential input. If you don't need to build a timegrid to monitor, maybe you can operate faster just by doing it by reaction, but it doesn't seem anything fancy to me at all.

>>11285589
I don't know if it is fancy or not but those methods are definitely not of public domain. I mean, we are talking about cutting edge custom hardware some of those firms need to constantly trade in the nanosecond domain. Machine learning is doomed if they are still working under regular clock timing conditions

>>11285661
A CPU runs on an internal clock, I think (not knowledgeable on hardware). Either they have analog system that places orders (big lol) or they still use some forms of cpus.

>>11285661
>>11285677
Also, the market maker also needs to process the order, that's a natural boundary for transaction speed from the point of view of the trader.

>>11285682
>>11285682
Order speed is probably defined by the market provider endpoint, but it's the decision time that matters the most for these operations. A building full of computers that work without clocks making insanely fast decisions is the crux of high frequency trading and sounds like the perfect setup for next gen machine learning architectures

>>11285696
HFT algorithms are incredible simple and are mostly people getting money out of the bid/ask spread when placing double orders

AI is just a way to keep the design of a software hidden behind the massive amount of data needed to create a classifier. no more 5 people leaving a company and starting another one and building a superior product thanks to the knowledge they obtained from their work, now you need knowledge of how to build a model + petabytes of data + thousands of indians being paid $20/10,000 images classified

>>11285750
>mostly people
>people
>human beings
>nanosecond processing speed computers are people

unironically ok boomer

>>11286029
You got the point, faggot. I obviously meant the methodology, not that it wasn't the the machines placing the orders.