Ask any AI hater to explain what's happening in this picture and watch them go quiet and then start repeating some nonsense marxist talking points they've memorized.
>>107676078AI haters BTFO once again
>>107676078can you explain it?
>>107676078>AI hatertrannies redditors and deviantart griftersthere is literally no other demographic hating AI
>>107676078...overfitting?
>>107676078log scale>>107676117always
>>107676117That's only this part of the graph. Why are you ignoring the rest?
>>107676078So what's your point here?
>>107676078first it didn't know then it learned
>>107676210I think he's trying to disprove 'it only learns patterns in training data' by showing that it can get 'smarter' even after it already figured out the training data
>>107676078I don't know whats happening in that picture.
>>107676112sooooooooo much this!!Jesus himself would be vibe preaching if he was still around
What the fuck happens here?
>>107676111ofc not. This is a homework thread.
>>107676078>picture of AI shit>why can't AI haters identify this?!?why would an AI hater be doing AI training?
Plebs vastly underestimate the amount of data intelligence requiresit's not the brute force approach is impossibleit's that you need a million times more data than its currently available
>>107676299Quantum entanglement
>>107676547would help if they didn't filter 99% of available data out for being "toxic" of some kind, ie not politically aligned with the lab.
>>107676078>hurr durr i just read deep double descentop is a nigger
>>107676078>generalizing a trivial function only needs a thousand TIMES more steps than learning the training data>still doesn't reach 100% of accuracy, but who needs your calculator to be able to calculateImpressive, very nice.
>>107676617you're supposed to call the tool to do math bro, it's for text not maths
>>107676547>it's that you need a million times more data than its currently availableThe higher-order the thing you're looking for, the more data current AI needs to find it. Not sure if it's quadratically more or exponentially more.But sure, let's not put any effort into finding a more efficient method.
>>107676111>can you explain it?It's called "grokking". You get 0 error on some artificial training objectives (modulo arithmetic here), if you train way past what you normally would.
>>107676078Looks like overfitting. You continue bruteforcing until you have high accuracy on the validation set, so even if the NN is never explicitly trained on it it still ends up overfitting it
>>107676112anyone have that reaction image from /pol/ of nazi pepe the frog saying jews are based that's how i feel about this new ai hate=tranny d/c bot spam
>>107676907fortunately for Sammy none of his investors have the 110 iq required to understand this post
>>107676078>That gapShould be on a /h/ image
>>107676907https://www.youtube.com/watch?v=D8GOeCFFby4
>>107676078Any real socialist/communist is pro AI, the luddites are just finding whatever political excuse they can to curtail progress
>>107677045Giving it a name does bot legitimize it. It's just bruteforcing and overfitting. Normal NN training has a loss measure that fits the training set. When you keep training after the model already fits the training set, until it fits the validation set, then you (or rather the "engineer" overseeing the training) become the loss runction. People just misunderstand statistics and neglect the bias inserter by their own actions. Choosing to cut training at a specific point or extend training further, based on the model's performance on the validation set, makes the validation set part of the training and taints the data. This is textbook overfitting.
>>107676078Another fucking>i watched a youtube video now i am smart let me make a 4chan thread about itthread
>>107677086I posted that video because a) it's probably what inspired OP to make this thread and b) it explains exactly how a model like this learns to generalize on the training set some time after the overfitting. It's an example of a highly interpretable network.
>>107677118It's not learning. If it already has 100% accuracy on the training set, but fails in the validation set, then it found a solution that fits all the training set but is not the correct one. By continuing to train it it tries different solutions that still fit the training set to 100%.How does it know it found the correct solution out of the multiple ones with 100% "accuracy" on the training set? It doesn't. The researcher looks at the validation set accuracy and decides whether the current solution is the right one or not. This is just training on the validation set by proxy.Eventually it will find a solution that fits both the training and the validation set. That's when the researcher stops the training. Does that mean it's the correct solution?No. Not necessarily. If you try a different validation set at that point, it quite likely will fail. Or if you continue training, it might go for a different solution that drops the validation set accuracy. It looks good in graphs, but in practice this is just blatant cheating. You are exposing the gradients to the validation set, making your "validation accuracy" meaninglessYou could get the same effect by simply not splitting the data into a training set and and validation set, simply train on all available data. It would converge faster and be just as smart (dumb)
