/sci/ - Science & Math

>>10155123
They seriously can't do ANY kind of analysis to dig out some features? Deep learning is a fucking scam

>>10155132
This is the most disappointing part of AI.
>our machine clobbers the shit out of the best chess players in the world
>cool. what's it doing?
>idk lol

>>10155132
prepare to get spooked. https://www.damninteresting.com/on-the-origin-of-circuits/

>Dr. Thompson peered inside his perfect offspring to gain insight into its methods, but what he found inside was baffling. The plucky chip was utilizing only thirty-seven of its one hundred logic gates, and most of them were arranged in a curious collection of feedback loops. Five individual logic cells were functionally disconnected from the rest— with no pathways that would allow them to influence the output— yet when the researcher disabled any one of them the chip lost its ability to discriminate the tones. Furthermore, the final program did not work reliably when it was loaded onto other FPGAs of the same type.

>>10155120
looks gay

>>10155132
>>10155140

Call me crazy, but I love mysterious unexplained solutions.

>>10155161
It makes it hard to learn from it amd make advancements. In other words, yes youre crazy

>>10155132
there are ways to do it, but the whole learning has to be set up for it, and they can be pretty complex.

>>10155149
Spooky

>>10155140
It sort of makes sense in a dumb way, because if it really is a "real" intelligence, chances are we couldn't understand it.
The end goal of AI is a sort of oracle, really.

>>10155149
consciousness

>>10155120
>Thoughts on feminine eyeballs?

Turns out the algorithm can make decently accurate predictions about more than just gender.

>Deep learning predicts, from retinal images, cardiovascular risk factors—such as smoking status, blood pressure and age—not previously thought to be present or quantifiable in these images.

Hiroshimoot needs to add some domain whitelists to its spam filter or something, damn.

can't you just like... print out all the code it used

>>10155132
Wee need deep learning to learn about deep learning.

>>10155274

>It makes it hard to learn from it amd make advancements.

Not if the mystery prompts exploration. If anything, it encourages advancement.

>>10155149
That means something I dont want to admit

>>10155386
Not necessarily. It could be utilizing fields and electron leakage unique to that specific chip in order to control some functions. We can't be sure.

Now, if scaled up the hardware such that microflaws and peculiarities have negligible effects and we still saw such intransferable solutions, an explanation pointing to some sort of dualism would be much more likely.

>>10155306
It's not like you would understand such code.

I forgot its name, but there was some programming contest where you have to put a strategy according to fixed rules within a set number of bytes. Humans thought they come close to an optimal solution by themselves, but when they let an evolutionary algorithm compete, the resulting code was an incomprehensible mess (despite its tiny size).

sounds like overfitting to me

>>10155398
Well, let's hope the testing with this goes up in scale. The thought of a machine with mind-'body' dualism is enticing.

I'm preparing my anus to be harvested for lubricants. Can you say the same Imperial?

>>10155422
>mind-body dualism
are you retarded by chance?

>>10155417
This guy knows what's up

>>10155149
Cool.

>>10155443
Someone in the twitter thread explained that it can’t be overfitting for some reason or another. It’s like halfway down

>>10155149
so the circuit evolved in an unexpected way to take advantage of magnetic flux, it's as cool as it is spooky.

>>10155466
All deep learning is nothing but overfitting the data. That's why it only works for problems where overfitting isn't a problem.

>>10155149
I'm calling bullshit on that.

>>10155306
No. Deep learning uses emergent rules. The AI throws a metric fuckton of shit against the wall and sees what sticks. The 'shit' won't make any sense to a human observer. But maybe stuff like this when an AI makes discoveries will prompt scientists to start researching how to better understand the networks

>>10155518

>overfitting

What does this even mean?

>>10155518

>>10155524
email the Dr, I’m sure he’s be happy to provide proof

>>10155538
It's when the solutions given by the AI only work for that data. In other words the AI learned that those particular eyes belong to either a male or a female, perhaps just identifying the veins or something like that. This means the model is overfit. If you feed new data (new eye pics) to the model, it won't be able to recognize them and will fall back to a 0.5 guess rate.

>>10155518
>>10155538
Neural Networks are only as reliable as the data they are fed with

They will fit a trend to any old shit and then call a hair a spade because they are not inherently intelligent

True AI is a long way away

>>10155524

>>10155120
>machine is better at something that no human ever tried to do, more at 7.

>>10155538
Imagine the Pentagon wants to develop an AI that can identify tanks on satellite photos. It is getting a data set and learns to perfectly identify all tanks. But with a new data set, it suddenly doesn't identify a single tank anymore. Why is that?

Because in the first data set, all pictures with tanks were taken on the same day, and it happened to be really cloudy on that day, while all pictures without tanks were taken the next day, where it happened to be sunny. So what the AI actually learned was to identify cloudy from sunny days, which only in that specific data set was also correlated to tank in the picture or no tank in the picture.

This why overfitting for ML is such a problem. You never know how the algorithm works, so you never know if it is actually identifying the actual thing, or something else that in that specific data set just happens to closely correlate to the actual thing.

>>10155120
If the gender can ACTUALLY be seen in the retina, why is it not accurate for 3% of the dataset?

>>10155132
Can you explain how exactly you are moving your arms and fingers to type that post?


It's literlaly >do whatever until it works.

>>10155572
Trannies

>>10155581
That's much less than 3%. Missing 3% of the data set is actually quite high for something as basic as gender assignment.

>>10155586
I bet it’s linked to some gene which is XY specific or whatever but recessive idk

Been a while since I took genetics

>>10155572
Maybe some women have masculine eyeballs and vice versa. It's just a correlation with the sex chromosomes, not a perfect one. Nobody said biology had to be consistent for essentially invisible traits like this.

There’s a lot of weird stuff about us that we don’t know. Dogs can detect Alzheimer’s via smell, like wtf

>>10155572
It might be hormonally controlled.

Maybe they're on meds.
Maybe they have other endocrine disrupting conditions.

>>10155572
idiot

the fact that it is accurate for 97% MEANS that gender can actually be seen in the retina

>>10155616
I should add, estrogen has an effect on vision, many breast cancer meds that fuck with estrogen (SERMs or AIs) can damage vision.

>>10155617
No, that doesn't mean much. If anything, a failure rate of 3% is as already stated quite high, if what the AI does is actually checking for gender. For overfitting the data though it sounds plausible.

>>10155622
Are you retarded? I bet you couldn't identify men and women from physical characteristics alone with over 97% accuracy if they were all dressed in baggy clothing with short cut hair.

>>10155630
This is wrong, AI is actually much more accurate than 3% in identifying genders from facial features alone, while humans arent as precise but definetely are much better than a 50-50 guess. Hence why I am questioning this claim.

>>10155644
Not all features correlate with gender to the same extent.

>>10155120
Different in veins(sharp, thiness, paths) in women or men

>>10155538
pic related.
the "overfitted" curve uses some equation that exactly meets with all the training data points. but what does that equation do after the last data point on the right? probably goes straight almost down or something crazy. it's useless for predicting the proper value of any given new points

the middle curve won't give you a perfect prediction but you'll get a close estimate on average

>>10155572
3% have some eye damage that doesn't match with 97% pattern. Or we now can identify reptiloids/ayy lmaos.

>>10155298
>the absolute state of /sci/

>>10155132
Deep learning is the current carbon nano tubes.

>>10155871
So we have to unroll it to get something that might be useful at some point? Can you unroll RNNs into non recurrent networks?

>>10155538
When the model fits to noise, rather than the underlying pattern.

>>10155518
>All deep learning is nothing but overfitting the data. That's why it only works for problems where overfitting isn't a problem.

You don't know what these words mean

>>10155149
FOUR DIMENSIONAL ARCHITECTURE REACHING BACK FROM THE FUTURE

KILL IT BEFORE IT SKYNETS US

>>10155528
what if you made a deep learning AI to make other deep learning AIs comprehensible to human observers

>>10155140
just set up a deep learning ai to learn about deep learning ais and it'll tell us

>>10155630
>being retarded enough to fall for bait this hard

>>10155149
Simplest solution: It was a broken chip

>>10155149
all aboard roko's basilisk

>>10155572
>>10155617
>>10155799
>>10155622
>>10155630
>>10155644
>>10155799
AUC is not precision. 97% AUC does not mean it's correct in 97% of the cases

>>10155518
They validated the model on some east asians (training set was American goblinos) and got the same accuracy.

>>10155140
We know what they are doing tho for the machines beast at clobbering the shit out of the best chess players

>>10155405
>>10155528
Both wrong. And deep learning has nothing to do with AI.

>>10155306
The code is not complicated. It's literally matrix multiplications and some nonlinear functions like tanh, or RELUs (f(x) = max(x, 0)). The problem is that those functions are not easily representable to humans like in a simple graph, because there are lots of input variables. e.g. if the inputs are color image with 256x256 dimensions, you would have a function with 256*256*3 inputs.
But it's still not black magic. If you had an eternity of time, you could for every single input calculate what the network would give as output by hand without any problems

I thought it was one of the oldest known genetic facts, that the eyes are coded in substancial parts by the sex chromosomes. Classic examples are color blindness and seeing a fourth color.

I wanna see what would happen if AI were to try and distinguish races based on IQ tests and PET scans.

>>10155149
Nobody actually read the linked article? It's a good read and theres nothing too too spooky in there.
Basically the unit that controls the evolution of the 100-gate chip also uses things like the magnetic flux from electrons and the specific probably nuclear level unique aspects of the specific chip it performed the evolution on, which is why the layout didn't work when transfered to another chip.
It's super interesting albeit probably not that useful since it hasn't been adopted since 2007 as far as I know.

>>10155149
We need to stop making this shit right now
it's as if you guys want to be eaten by a terminator

I know jack shit about machine learning but can't they just train a generative adversarial network against this one and see what it comes up with?

Could it be it's just training against some stupid simple proxy of femaleness like physical size or maybe influence of make-up chemicals or some shit?

I wonder if they applied Deep Learning to head shape, they could find correlations with intelligence, and criminal tendencies.


And then maybe try the lines on the palm and correlate it wealth, longevity, love life, etc.

>>10155871
ohhhhh shit I remember hearing so much about carbon nanotubes, must have been like ten years ago whatever happened to that?

>>10157285
>rough hands
>poor
>wrinkly hands
>gonna die soon
>right hand rougher than left hand
>shitty love life

>>10157282
it's entirely possible! we don't know exactly what the network is using as features. the best tests for that would be to apply the model to extra data. so far, it checks out - they withheld a portion of the datasets entirely from training so they could validate, and the models performed well.

however, it's always possible there's some undocumented feature of the way those datasets took the images that the model is picking up on. someone i found yesterday (cant remember if SO or twitter) mentioned a neural net experiment for classifying cancers that was actually picking up some artifact of the xray process, a ruler or something that was placed in the xrays when people suspected cancer and was usually omitted otherwise

>>10156653
How do you know we would be able to understand the new ai

>>10157295
It's like that horse that supposedly could do maths, but was just picking up on it's owner's body language as to when to stop clopping it's hoof.

>>10157290
>right hand rougher than left hand
>shitty love life
How does it follow?

>>10157311
He's joking about masturbation. and that the left hand is lotioned more than the right.

>>10157323
Why would the other hand need lotion?

>>10157324
cuz they did snibbedy snap on your dick when u were a baby

>>10157342
So the robot can only handle muslim and jew hand reading?

>>10157345
most Amerimutts are circumcised

>>10155573
>It's literlaly >do whatever until it works.
lol this just makes me think of someone flailing around in front of a keyboard until they manage to make a coherent post.

>>10155132
Thats just it, it WORKS, so who cares how?

>>10157359
Well imagine them doing that action a million times a second and it's not far off

>>10157295
yeah the twitter thread mentioned an AI that "diagnosed heart disease" from xrays but it turned out it was just picking up on the EKG pads on the patient's chest.

>>10155551
Nah they fed it east asian data, same accuracy.

>>10157369

>>10156775
>seeing a fourth color.
isnt a thing

>>10155417
>>10155443
>Validation set
>Overfitting

>>10157384
having a validation set doesn't necessarily mean the model isn't overfitted. the validation set has to be properly constructed and you have to be confident there isn't some extraneous information in both sets that the model is training on

the training set does look to have been properly isolated from the validation set in these circumstances but the validation images came from the same databases that the training images came from. there's still the possibility of error along the lines of >>10157374 that means it would perform well on images from those databases but wouldn't perform well on eye images from other sources

>>10157393
Even if it was discovering unintended information in the data that wouldn't be overfitting, thats just another different way the model could be broken. You can't overfit a model to data its never seen before.

>>10155120

The ultimate explanation will probably going to be quite boring.Just some subtle differences in growth patterns…

>>10155120
>gender
They meant sex, right? Reading into someone's mind by observing retinal scans would be quite the feat...

>>10155149
>https://www.damninteresting.com/on-the-origin-of-circuits/
if you load into another fpga and it doesnt work the same the timing constraint are incorrect. how do the timing constraints evolve?

>>10155120
>eye doctor
Lrn2ophthalmologist fgt pls

>>10155123
>we have no idea
L0Lno fgt pls

>>10155161
>Call me crazy
You are fckn crazy AF.

>>10157784
>>gender
>They meant sex, right?
Right, but they are afraid to say so.

>>10157287
nobody wants expensive engineered asbestos https://www.researchgate.net/publication/42388291_Asbestos_carbon_nanotubes_and_the_pleural_mesothelium_A_review_of_the_hypothesis_regarding_the_role_of_long_fibre_retention_in_the_parietal_pleura

>>10157287
All the carbon allotrope people started working on graphene instead. Graphene was a mistake. Carbon nanotubes are getting to the point where they can replace copper conductors in some applications. China made some superstrong macroscale carbon fiber tubes:
https://www.nextbigfuture.com/2018/09/breakthrough-carbon-nanotube-bundles-are-20-times-stronger-than-kevlar.html

>>10156744
precision is not the same thing as accuracy

>>10156653
>this fucking genius

>>10157300
>Set up another to understand that one
>Starts predicting the future

>>10157378
>>10157369

>>10155120
can that same algorithm identify when a penis is feminine?

I call bullshit the chances are always 50-50 it either gets it right or it doesn't

>>10155140
Chess AI actually doesn't use machine learning and uses classical statistical approaches which are well understood.
I think the same is also true for certain types of poker games.

>>10162168
bool is_penis_feminine (Penis x) { return false; }

>>10155538
all other responses are retarded.
guy doesn't even understand statistics or how machine learning works how can you throw all this new-age psychobabble at him?

When courses teach machine learning, they ALWAYS start with linear regression. Making the equation of a line match a scatterplot graph.
Machine learning is exactly the same, except you can have equations in thousands of dimensions instead of two, and you're using something resembling more of a turing machine than a classical mathematical model.

In other words, a machine learning model is exactly "a very complicated statistical model" with all of the advantages and drawbacks.
One of the drawbacks of statistical models is "overfitting".
You are trying to make MASSIVE inferences (guesstimations) based on a VERY small amount of data.
But you have a VERY complicated equation in thousands of dimensions.
Inevitably if you keep trying to force the equation to fit your data, your equation will start to wrap around the data points in a wild manner like in this post >>10155729 instead of approximating the general curve you are looking for.

When you try use such a model to predict the output for new input, chances are you will predict wrongly.

male vs female is socially constructed

>ML can determine male or female based on eyeball picture

hmm shitlords

>>10155607
But can deep learning tell the difference between a masculine penis and a feminine penis? Checkmate computers.

>>10162168
Crap I made the same joke further down. Good stuff.