javascript machine-learning neural-networks neural-networks-from-scratch typescript

Micrograd TS

This is a TypeScript version of karpathy/micrograd repo.
A tiny scalar-valued autograd engine and a neural net on top of it (~200 lines of TS code).

This repo might be useful for those who want to get a basic understanding of how neural networks work, using a TypeScript environment for experimentation.

Project structure

micrograd/ — this folder is the core/purpose of the repo
- engine.ts — the scalar Value class that supports basic math operations like add, sub, div, mul, pow, exp, tanh and has a backward() method that calculates a derivative of the expression, which is required for back-propagation flow.
- nn.ts — the Neuron, Layer, and MLP (multi-layer perceptron) classes that implement a neural network on top of the differentiable scalar Values.
demo/ - demo React application to experiment with the micrograd code
- src/demos/ - several playgrounds where you can experiment with the Neuron, Layer, and MLP classes.

Micrograd

See the 🎬 The spelled-out intro to neural networks and back-propagation: building micrograd YouTube video for the detailed explanation of how neural networks and back propagation work. The video also explains in detail what the Neuron, Layer, MLP, and Value classes do.

Briefly, the Value class allows you to build a computation graph for some expression that consists of scalar values.

Here is an example of how the computation graph for the a * b + c expression looks like:

Based on the Value class we can build a Neuron expression X * W + b. Here we're simulating a dot-product of matrix X (input features) and matrix W (neuron weights):

Out of Neurons, we can build the MLP network class that consists of several Layers of Neurons. The computation graph in this case may look a bit complex to be displayed here, but a simplified version might look like this:

The main idea is that the computation graphs above "know" how to do automatic back propagation (in other words, how to calculate derivatives). This allows us to train the MLP network for several epochs and adjust the network weights in a way that reduces the ultimate loss: