/sci/ - Science & Math

>>9743880
Only neural networks can contain a soul needed for the AI to be self aware. It simply doesn't work with standard programming. This is why dogs don't have souls.

Not science or math

>>>/g/

>>9743895
I was going to post there but I thought this was science and so I posted it here.

>>9743886
Go to /x/ brainlet, unless you’re going to prove souls exist in the lab.

>>9743880
neural networks can learn. If you want to use control statements you can only make a program as good as you are at that task. A neural net is self-correcting. You don't necessarily need that great of an understanding of the task, your neural net AI will figure it out on its own, given a good utility function.

Specifying a specific set of control statements to identify something like a handwritten letter is really hard (different handwriting styles, cursive vs. print, capital or lowercase, relative size/position on the page, etc). Rather than go through the agony of personally constructing and debugging some control sequence, someone realized that we can instead just use statistics to guide us through an algorithmic process of improving our guesses. Artificial neural networks happen to be a convenient system of writing down optimization problem, since a lot of linear algebra tools carry over (we’re optimizing the cost function “out of how many pictures of a handwritten letter did I guess the right letter”). Beyond that general idea, there’s a lot of complexity with how to not get stuck in a local minimum, how to not overfit the training data set, convergence to a minumum, etc.

Imagine the amount of conditional statements required to answer something like how many jaguars are any given photograph.

if you use control statements you need at least one for every single use-case, that can be a lot of work if not automated.

think of neural networks to automate this process meaning it can create it's own control statements dynamically by using memory and optimization, they might not be as consistent and precise as manmade control statements but they are very fast and efficient.

>>9743971
also this.

>>9743880
there are many algorithms. neural networks is just one of them https://en.wikipedia.org/wiki/Statistical_classification#Algorithms

>>9743971
>>9744758
not to mention that if your problem/input slightly changes then the abstracted network model/method can be retrained on these new inputs without having to laboriously modify each of the control statements or conditions.

>>9743880

Neural networks can be trained using backpropagation. How would you train a model that uses programming statements?

>>9743880

>>9744773
that post made me slightly anxious for a moment when I thought the "this" in the first line referred to the book

>>9744764
exactly.

Though, I thought my use of the word 'efficient' implied this clearly.

>>9743886
What is the cut off point for souls?

A: Homo habilis
B: Homo erectus
C: Homo Neanderthalensis
D: Homo sapiens

>>9744803
Homo Rufus

Neural networks allow AIs to make decisions by examining information, drawing conclusions, weighing options against goals and risks, and then deciding.

If-then trees are hardcoded.

>"but isn't a neural network just a super complex if-then tree"
Yes, and the human brain is a chemical reaction so complex it became self-aware.

The key to AI is don't overthink it.

>>9743880
Machine learning is 99% memes.

It's just statistics where you automate updating your model. It's not interesting and its applications are likely to be limited.

>>9745048
>Yes, and the human brain is a chemical reaction so complex it became self-aware.

Reductionism is the drug of retards.

>>9744773
Also highly recommend this textbook. This guy's a genius. Used continental philosophy to argue "AI retards" (as I prefer to name them) were being stupid and overly optimistic, and was proven absolutely 100% right by history.

It's worth remembering in light of all the "AI retardation" the history of AI research is filled with people promising next gen AI tomorrow and massively under-delivering.

>>9745067
>continental philosophy
>not analytical
ok, into the trash it goes.

>>9745077
Your biases blind you and mislead you.

>>9745077
Both are good

>>9743880
Brains are reflexive and self-referential. If-then statements would need similarly non-deterministic, and self referential inputs and outputs to replicate a brain's behavior.

>>9745080
>>9745092
it's shit, it's not direct, it's ambiguous, full of linguistic confusion, that's why mathematics and formal languages are good desu.

>>9745112
Math is full of ambiguous, shitty writing. It's par for the course.

Analytic philosophy isn't clearer at all, either. Ever try reading Wittgenstein?

>>9745268
>Math is full of ambiguous, shitty writing. It's par for the course.
it isn't

>>9743886
why are shitskins literally always brainlets.

>>9745335
fuck off, /pol/tard