One neuron, by the numbers.
Inputs 0.2, 0.5, 0.43 times weights 0.4, 0.26, 0.3 give 0.08 + 0.13 + 0.129. Add bias 0.1: z = 0.439.
Activation: σ(z) = 1/(1 + e⁻ᶻ), so σ(0.439) ≈ 0.608. That number is passed to every neuron in the next layer.
Training picks the weights w that minimise the average loss (1/n) Σ ℓ(h(xᵢ; w), yᵢ), by stepping down the loss surface along −∇L.
Learning is weighted sums, a squashing function, and gradient descent.
显示更多