Here is a link to my submission: https://neuralnet-visualization--samdevz.repl.co/
You should definitely watch it in a large window. It's not optimized for small screens yet!
So, what is it?
I created a Neural Network that can predict how many posts someone has made on their repl.it account using an open source library I helped build!
But that's not all! I wanted to make it engaging, so I created a visual element as well where you can:
Watch the neural network train (the thickness of the lines represents the strength of the connections between nodes)
Watch gradient descent take place (the graph on the right shows the loss (hopefully!!!) decreasing in real time)
Track the steps, validation accuracy, and much more!
To make its predictions, the network uses the following details about someone's profile:
Do they have a default avatar (yes/no)?
How many points do they have?
How many comments do they have?
It's very interesting watching the network learn and seeing what connections are important, and one ones ultimately aren't!
There is a lot more I want to add, but I also did not want to miss the deadline! I hope that right now it can still serve as a fun way to understand how Neural Networks learn and the code -- although a bit messy -- is actually fairly organized (definitely check out nn.js)!
This is truly amazing! I'm suprised more people haven't seen it. And it has a nice design as well