So here it is, one of those AI thingies that are supposedly going to take over the world any time now. Granted, this project will probably not help reach singularity. It's only reason for being, is really my sudden urge to create.
It is a deep neural network implemented in JS.
- XOR: The hello world of neural networks
- Digits: Classify handwritten digits using a subset of the MNIST database
- Fully connected layers
- Multiple activation functions (one per layer)
- Epochs and batches
- Progress callbacks
- Save and restore from JSON
- Basic visualizations
- Async training and testing
In short, the basics are here for playing around with deep learning on small problems.
The core functionality has no external dependencies.
This project is educational and driven purely by curiosity. It is not intended to compete with the myriad of implementations out there.
Potential fun for the future:
- SoftMax, OneHot, Cross Entropy Loss
- Convolutional layers
- Hardware acceleration
- Recursive layers
- Streaming input
- Distributed training