Neural Network Visualization

That’s a CNN, my prof invented it! Super cool guy

Source, btw:

This is incredibly succinct.

The raw information gets processed into much smaller mental category objects used for judgement and reaction - food, not food, threat, not threat, mate, not mate

Seeing the layers light up, the connections firing, and the whole thing pulsing like a living system makes the black box magic of AI suddenly feel tangible and beautiful.

can we please have bad apple in this animation...

That's an interesting math behind it, always surprises me

Second layer:

Input layer, two dense layers, and an output layer?

Can you draw a rabbit

这个神经网络3D可视化把从像素输入到特征提取、再到输出分类的整个过程展现得超级直观👍

What's the algo running behind this, just curious

The visualized version needs sorting on the second and 3rd level for the visualization to make sense. Without sorting: "Oh, a bunch of random things blip up when I draw, and the correct number comes out for no reason"

it’s just gradient descent with extra steps.

Neural networks are so fucking bad.

784 neurons in that second layer just went full disco

So cool. 😎👍🏽

Very compelling and intuitive

Very good

If only governance wouldn’t have killed tech… we could have been way further with tech. Genuinely help people… instead of this. Yeah I know I’m 🇳🇱;Smells like VoC… Also tech CEO 🤓. Pro #humanity #truth

Nice! I'd like to get this tool.

..best illustration yet!

Matrices o.o

It's just that the step of convoluting with some kernel literally makes zero sense. The max pooling layer seems to make sense even though it's confusing how afterwards it produces a likelihood.

That's something.....

Once upon a time

I did something similar some years ago with OpenGL and C++. MNIST Convolutional Neural Network implementation in C++

Cool! And you have retard Dawkins over here, "It's sentient!"

watching this, I am about to shed a tear bcz I remembered how I learned about CNN :) but now, OCRs are mostly transformers, especially PaddleOCRs. And trends indicate that OCRs are moving to VLM-OCR architecture

Nice