Weights, bias, activation

A weight scales an input before it enters the sum. For a positive input, a positive weight contributes positively and a negative weight contributes negatively; a negative input reverses those signs.

The bias shifts the whole output up or down regardless of the inputs — it sets the neuron's default leaning.

The activation transforms the weighted sum. A sigmoid maps it into the range 0 to 1; other activations behave differently. ReLU, for example, is zero for negative inputs and has no upper bound. A value between 0 and 1 is not automatically a calibrated probability.