This is a tool to experiment basic neural network architectures and build intuitions on machine learning.
The main idea is provide an environnement to experiment how data and learning parameters affects the network training.
The online demo is available here :
https://akarzazi.github.io/neural-network-playground
To get a first sight, click on the play button and watch the prediction evolving towards the labeled data.
The lab has two main sections :
-
Generate labeled data.
-
Train the network and watch the metrics.
Training data has the following shape : {a, b} => label, for a pair of numbers a and b, we will associate an expected number label.
a and b are represented on a 2d plane on the x and y axes respectively and the label is viewed by a color scale.
The goal of the network is to understand the relationship between {a, b} and the expected label.
The achieve this task you may need to tweak several things, the most important are :
- The neural network shape and activation functions.
- The batch size.
- The optimizer settings.
This project was inspired by great resources on the internet :
https://playground.tensorflow.org/
https://ruder.io/optimizing-gradient-descent/
https://www.youtube.com/user/shiffman/videos (the coding train)
We're glad to know you're interested in the project.
Your contributions are welcome !
You can contribute in the following ways :
- Report an issue / Suggest a feature.
- Create a pull request.