One more adjustable number

A weight belongs to one connection. It controls how much one input counts.

A neuron can also have another adjustable number. Add it after combining the weighted inputs, just before the neuron checks its result.

Suppose the neuron needs 0.8 before it responds, and the weighted inputs total 0.6. It falls short and stays quiet. Add +0.4 and the total reaches 1.0, which clears the bar. The inputs never changed.

That number is called a . It arrives before any input does, so the neuron starts out pre-filled and needs less evidence to respond. A negative one does the reverse, leaving the neuron to make up ground before it will answer at all.

Without a bias, a neuron could only change how much each input counts, never how much it takes to convince it.

Weights and biases are together called . A model described as having 50 million parameters contains 50 million such values. Training may adjust all of them, or keep some frozen while it updates the rest.