NumPy-only ML zoo: 30+ algorithms, zero black boxes

What it does

ML-From-Scratch is a collection of fundamental machine learning algorithms implemented in pure Python with NumPy. The repo covers supervised learning, unsupervised learning, reinforcement learning, and deep learning — everything from linear regression and SVMs to CNNs, GANs, and Deep Q-Networks.

The interesting bit

The author explicitly trades speed for transparency: no PyTorch, no TensorFlow, no opaque autograd. You get hand-rolled backprop, evolutionary neural networks trained by mutation and crossover, and even a genetic algorithm that evolves strings letter by letter. It is pedagogy through masochism, and it works.

Key highlights

• 30+ implementations including XGBoost, DBSCAN, Restricted Boltzmann Machines, and Apriori association analysis
• Each example prints a model summary showing layer shapes and parameter counts, like a tiny Keras built from np.dot
• Working demos: polynomial regression on Swedish weather data, CartPole via DQN, MNIST generation with a GAN
• Includes neuroevolution and particle swarm optimization of neural networks — gradient-free training methods rarely seen in introductory material

Caveats

• The README warns that efficiency is explicitly not the goal; these are reference implementations, not production code
• Some README sections are truncated in the source, so the full list of unsupervised and deep learning implementations is cut off

Verdict

Ideal for students, interview preppers, or anyone who has called .fit() a thousand times and finally wants to see the matrix math. Skip it if you need GPU acceleration or a library to ship tomorrow.