I Built a Neural Network from Scratch in SCRATCH

A developer builds the same neural network for handwritten digit recognition (0s and 1s) in four progressively difficult implementations: PyTorch, NumPy, pure C, and finally Scratch (the visual programming language). The journey demonstrates fundamental ML concepts like forward/backward passes, loss functions, and gradient descent, starting with high-level abstractions and ending with manual matrix operations using logic blocks. The Scratch implementation required downsizing images from 28x28 to 10x10 pixels and reducing hidden layer neurons to 10, but successfully trains and classifies digits.