Brain's Decision Committee
A 10-part notebook series that teaches neural networks from scratch through one visual problem: classifying vertical and horizontal lines with NumPy, matplotlib, and interactive experiments.
Start lesson oneWhat you will build
- A V/H line classifier from scratch
- A multi-layer neural network class
- Training, evaluation, saliency, and troubleshooting loops
Lesson path
Ten notebooks, normalized into an AgenticWorks player.
The language of the committee
MatricesMeet the Brain's Decision Committee and learn why images become matrices before a model can reason about them.
The first committee member
Neuron anatomyBuild a single artificial neuron and watch inputs, weights, bias, and weighted sums become a first vote.
The vote
NonlinearityCompare Step, Sigmoid, Tanh, ReLU, Leaky ReLU, and Softmax as different voting methods.
The first prediction
Forward passTurn the neuron into a perceptron, generate a V/H line dataset, and make first predictions.
Learning from mistakes
BackpropagationMeasure error, apply gradient descent, tune learning rate, and implement the training loop.
The trained expert
MetricsSeparate training from inference, compute evaluation metrics, and inspect what the model learned.
The full committee
MLPMove beyond one neuron with hidden layers, XOR, and a multi-layer perceptron.
Growing pains
Failure modesSee overfitting, vanishing gradients, exploding gradients, and the common fixes.
Mastery
End to endAssemble the complete neural network class, data pipeline, training loop, evaluation, and dashboard.
Where the road leads
Future pathsConnect the from-scratch model to CNNs, RNNs, Transformers, frameworks, and projects to build next.