Un video YouTube.
Sei formati pronti da pubblicare,
in 2 minuti.

Executive summary — The video explains, with no intimidating equations, what a neural network is by starting from a concrete problem: recognizing a handwritten digit. A neuron = a value (its "activation"). Neurons are organized in layers: the input layer reads the image's 784 pixels, hidden layers combine patterns (strokes, loops), the output layer yields the digit. Each connection carries a weight, each neuron a bias; an activation function (sigmoid) squashes the result between 0 and 1. The key idea: the network learns to detect sub-patterns and assemble them hierarchically.

Takeaway: a neural network is just a large function parameterized by thousands of weights and biases — learning means tuning those parameters.

So what is a network, really?

A neural network starts from a simple question: how can a program recognize a hand-scrawled 3 when no two 3s ever truly match? The answer is one word: hierarchy.

Neurons and layers

Each neuron holds a number between 0 and 1 — its activation. The first layer takes the 784 pixels (28×28). Hidden layers combine those signals to spot strokes, then loops, and finally whole digits.

Weights, biases, activation

Every connection has a weight; every neuron a bias. The weighted sum passes through a sigmoid that squeezes it between 0 and 1.

Learning means tuning those thousands of weights so the right output lights up.