Watch: Neural Network
Layers of weighted sums passed through a non-linear function, with the weights fitted by gradient descent. The base structure under every current AI model.
Transcript
At the heart of modern artificial intelligence is the neural network. It is a structure made of stacked layers that processes information step by step. In each layer, the network takes an input, multiplies it by a set of weights, adds a bias, and passes the result through a non-linear activation function. Stacking these layers is what makes the network expressive enough to understand language, images, and audio.
That non-linear step is crucial. Without it, stacking multiple layers would be pointless, collapsing the entire system into a single, simple layer. The non-linearity is what allows a deep network to bend and shape the input space, making complex patterns separable.
The weights and biases are the model's parameters. They start out completely random and are gradually tuned using backpropagation against a loss function. Once training is complete, these parameters are frozen into a checkpoint. This is the model's parametric memory, and it is where everything the AI knows is stored.