Bias (parameter)
An essential learnable parameter added to the weighted sum of inputs in a neural network node, allowing the activation function to simply shift left or right.
Think of It Like This
Like adjusting the baseline temperature of an oven before you decide how much you want to turn the heat up or down based on the recipe.
Without a bias term, a neuron's activation boundary is strictly forced to pass through the origin of the input space, severely limiting its flexibility. It acts exactly like the y-intercept in the classic linear equation y = mx + b. By adjusting the bias, the network learns to activate even when all input features are zero.