Backpropagation by Hand

Most explanations of backpropagation skip the part that actually builds understanding. They give you the high-level story (chain rule, gradients, update rule) and then hand you a library that does the work. You leave feeling like you understand it, but you couldn't reproduce a single gradient if asked. This course takes the opposite approach. You start with a tiny network, specific numbers for weights and inputs, and you compute everything yourself: the forward pass, the loss, the output-layer delta, the hidden-layer gradients, and the weight updates.

The goal is not to avoid using libraries. The goal is to get to a point where you know exactly what those libraries are doing. When a network fails to train, the people who debug it fastest are the ones who can reason about what the gradients must look like at each layer. That skill comes from having done the calculation yourself, slowly, with a pencil, at least once. This course gives you that experience across four progressively harder lessons, ending with a two-example mini-batch and a review of the mistakes that trip people up most often.

You do not need a GPU, a Python environment, or a deep math background. You need the ability to multiply and add, some patience with notation, and a willingness to work through the same small network from multiple angles. By the end, you will have run a complete training step by hand and watched the loss go down. That moment changes how you see the algorithm.