NN

Simple autograd engine implementing backpropagation and a neural network library built on it. Heavily inspired by micrograd and PyTorch autograd.

The main.rs file shows an example of MLP neural network performing binary classification on the make_moons dataset. The model is trained on 100 points of the make_moons dataset. After training, it is tested on the remaining data and achieves 100% accuracy.

Here's a visualization of how the decision boundary evolves during gradient descent and the correctly classified data points.

Run

cargo run

> cargo run
Training set gradient descent
Step 0, loss: 1.6074702669917762, accuracy: 54%
Step 1, loss: 1.7333523672032618, accuracy: 47%
...
Step 98, loss: 0.011572048264185729, accuracy: 100%
Step 99, loss: 0.011518733035234319, accuracy: 100%
---------------------------------------------------
Test set inference
Accuracy: 100%

Output plots

cargo run -- --plot && magick -delay 40 -loop 0 decision-boundary-*.png decision-boundary.gif