Learn ML Algorithms from Scratch

The problem

Libraries make it easy to train a model without seeing what an optimizer, a loss or an activation actually computes. That leaves beginners unsure which piece to change when a model does not learn.

The approach

I wrote each component in NumPy alone, as small modules with one job each. Linear regression is built from them end to end, with worked examples.

• Gradient descent, Momentum, RMSProp and Adam optimizers.
• MSE, MAE, Huber and cross-entropy losses, and sigmoid, ReLU, tanh and softmax activations.
• Accuracy, precision, recall, F1 and R² metrics, with scalers, encoders and a train/test splitter.

The result

An open-source study project with a test suite for every module, which pairs with Neural Network from Scratch and LSTM Gold from Scratch.