Grad School: Build a Neural Network Framework from Scratchhttps://indicogni.com/build-your-own/neural-network/2026-09-30T00:00:00ZRD01 A function with knobshttps://indicogni.com/build-your-own/neural-network/lessons/01-machine-learning-fitting-a-line/2026-09-30T00:00:00ZMachine learning starts with a function that has knobs. Fit a line to taxi fares by eye with two sliders, then write predict_line in plain Python.02 How wrong is wrong?https://indicogni.com/build-your-own/neural-network/lessons/02-mean-squared-error/2026-09-30T00:00:00ZWhy a loss squares its errors: signed errors cancel. Write mean squared error, mean absolute error and a grid search over lines in plain Python.03 Downhill in the darkhttps://indicogni.com/build-your-own/neural-network/lessons/03-gradient-descent/2026-09-30T00:00:00ZGradient descent from scratch: measure a slope by nudging, step downhill, and see why too big a learning rate never arrives. Plain Python.04 The simplest decisionhttps://indicogni.com/build-your-own/neural-network/lessons/04-perceptron-sigmoid-neuron/2026-09-30T00:00:00ZA step function can't be trained: its slope is zero. Build a sigmoid neuron in Python, train it on AND and OR, and watch it fail on XOR.05 Bending spacehttps://indicogni.com/build-your-own/neural-network/lessons/05-nonlinearity-xor/2026-09-30T00:00:00ZTwo linear layers collapse into one. See why, add a ReLU kink, and compute XOR with a two-layer network written in plain Python.06 The bill for nudginghttps://indicogni.com/build-your-own/neural-network/lessons/06-finite-differences-cost/2026-09-30T00:00:00ZFinite differences cost one forward pass per parameter. Count the calls, plot the bill, and see why 10,000 knobs need backpropagation.07 A number that remembershttps://indicogni.com/build-your-own/neural-network/lessons/07-computation-graph-value/2026-09-30T00:00:00ZBuild a micrograd-style Value class in Python that records its parents and its operation, and draw the computation graph it makes.08 Who owes the gradient?https://indicogni.com/build-your-own/neural-network/lessons/08-backpropagation-chain-rule/2026-09-30T00:00:00ZBackpropagation is the chain rule: local slopes multiply. Work a small graph backward by hand, then give + and * their backward in Python.09 Many paths, one sumhttps://indicogni.com/build-your-own/neural-network/lessons/09-gradient-accumulation/2026-09-30T00:00:00ZA value used twice gets gradient from both paths. See how = instead of += silently breaks backprop, and fix it with gradient accumulation.10 The order of blamehttps://indicogni.com/build-your-own/neural-network/lessons/10-topological-sort-backprop/2026-09-30T00:00:00ZBackprop must visit the graph in topological order. Watch breadth-first get it wrong, then write an iterative topological sort and backward().11 Trust, but verifyhttps://indicogni.com/build-your-own/neural-network/lessons/11-gradient-checking/2026-09-30T00:00:00ZGradient checking with central differences and relative error: measure five derivatives, catch a planted wrong formula, and write gradcheck.12 A network that trains itselfhttps://indicogni.com/build-your-own/neural-network/lessons/12-neural-network-from-scratch/2026-09-30T00:00:00ZBuild a neural network from scratch: Neuron, Layer and MLP on your own autograd engine, trained on XOR and two-moons in your browser.