Neural Network 3D Simulation 

thanks

On an unrelated note, how do I give a comment -1 likes? Asking for a friend.

I agree

@@DenisDmitrievDeepRobotics ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-0U_FBHKYqRk.html

I only love neural nets

@David Parry there's something going on inside of the head of a stereotypical stereotyper... unfortunately, you rarely get to see their regression model. oh well. who am I to judge?

​@David Parry we may not have anything going on up there, but at least we're finally as smart as you.

Beautiful comment

Inside out? Where did I see this dialogue??

@@kinseysharon2672 inside out yes

@hyper Computational neural networks model brains the same way a stick figure models a person. 웃 Adding a couple of lines for hair, might might the stick figure more closely model a person but it's still a bad model. The interconnected nature of brain neurons might be partially modeled by neural networks but there is a lot of brain structure that is left out. It's not just a matter of scale, it's a matter of kind. For example, hormones are integral to brain function but neural nets have nothing that is analogous.

@@myothersoul1953 Hmmmm... All models are wrong, some are useful. Just because it is incomplete and doesn't get anywhere near the scale or complexity of the brain doesn't make it a bad model. What's important is to know what you are trying to learn from the model and if it serves that purpose it is a good model. Got a better idea for computationally modelling the brain?

@@TheRealJavahead That it doesn't get near the scale of complexity or include major components of the brain makes it a very poor model. Just like a stick figure is very poor model of human anatomy. Fortunately I am not in the business of modeling the brain because that's a very challenging task. Before a model could build how the brain functions needs to me understood. Many scholars across the world are working on parts of that. I'm sure they would tell they are making progress but are we no where near a complete understanding. You are right, what you are trying to learn from the model is important to know. What aspects of the brain you model focuses will vary depending on whether you are interested on perception, cognition, emotions or motor control. It would be a huge mistake to think the such a simple model such as neural nets come close to representing any of those.

@@myothersoul1953 I do note that you have gone from categorising this as "a bad model" to "a poor model," next step might be that this is an "acceptable model." ;-) Relying on a reductio ad absurdum argument with your stick figure analogy is a false equivalency and doesn't support your assertion that this is a very poor (or bad) model. My point is, neither of us is in a position to determine how bad or good the model is without knowing its purpose.

@@TheRealJavahead The purpose of a model is to represent something. A poor model does a bad job of representing. The reason I use the stick figure analogy is neural nets are often represent with lines a circles which are the same things used to draw stick figures. Stick figures only represent actual humans at the grosses of levels and neural nets resemble brain function at the grosses of levels (at best). That is neural nest are poor modes. But you are right, all analogies break down when pushed. The neural net analogy of the brain just needs the slightest breezy to fall down. A better representation of neural nets would be a set of equations and procedures but that still doesn't match the brain. That's why they do such a bad job of representing how brains actually function.

it has already been done by deepmind, they used neural network to tune hyperparameters of other neural network.

Already done. Check out Google's AutoML !!

Ah, but where is the neural network that tests neural networks on their ability to discover better neural networks?

@@J4hk2 you can always go deeper.

check out HyperNEAT

Ben Kahrmann what? Do you even know anything about machine learning?

@@w花b It's not.

@@w花b it's 50% slower

タメル Who exactly... said that?

@@w花b lol

@DarkXSeries7 how about you can easily code it ;) in python or c#

They are super difficult to train. But they have been shown to be potentially more powerful than any ANN

@@kronek88 I think it depends on the problem you want to tackle, if you need a immediate approximation of something that requires a lot of computational power, then yes it's useful. But if you want a very precise result of a problem that isn't time consuming computationally. Then I think there are better algorithms than this one.

@MetraMan09 Yes, they mimic natural human brains. Like they can try and solve a task, approach it another way, check their answer, etc. It's kind of able to divide itsself up and basically train another couple neural nets to each do their thing to work together to solve complex tasks. It requires a LOT of computational resources but also can solve tasks static networks simply can not solve. It is, however, capable of being potentially dangerous if not done correctly. In theory, this kind of network is the kind of artificial intelligence that we theorize could pass for human in a turing test.

Also spiking neural network could create autonomous AI system, simulate memory and brain functions.

How? Can you elaborate?​@@kotcraftchannelukraine6118

pls make 3d visulaization of that lol

Суть всей "магии" заключается в том, что скрытый слой можно настроить так, что на любое входное значение, на выходе будет соответствующий нужный правильный выходной сигнал, который мы хотели получить.

It is not the image that moves - it is our brains automl experimenting with the bias hyperparameter.

Computational Neuroscience....that seem awesome

Tutorial how to visualize neural network would be nice. This video is in my recommendation, and it's very mind blowing. Do you think this can help to debug NN? For example adjusting dropout rate? I will be very happy to follow your progress on this kind of video, is there any link, or github perhaps?

@@ErlandDevona this video in particular is not intended to help you in droprate adjustments. Probably will cover droprate effect on next video, also will consider putting source code into github. Thanks.

@@DenisDmitrievDeepRobotics Hi. Did you put the source code to the github or somewhere since then?

Have a look at Sebastian Lague on RU-vid. He did a video visualising what the weights and biases do.

I feel you

Aj Ulyate Thanks for your observation.

I smell... overfitting

@@trentonpaul6376 Depends on the size of the dataset.

3? I guess that's a lot... If you're training on like an embedded graphics chip 😆

@@NilesBlackX 3 is for really non linear data. Written numbers doesn't need more than 1 hidden layer

@@terrykarekarem9180 MNIST doesn't, but proper handwriting training benefits from more layers. Adding punctuation, capitalization and diacritics makes the complexity rise above O(26) by an exponential. Regardless, 3 hidden layers isn't 'a lot' by any means, since that's a relative term which implies greater than average, and average depth for a FFNN definitely isn't < 4.

Next video is coming soon.

@@DenisDmitrievDeepRobotics great news!

thanks

Namaste 🙏

And the second song?

i do not know

It's called namaste indeed, from audiotronix