Why Filters Have One Bias Value
Problem
Why does each filter have only one bias value instead of giving each filter entry its own bias?
Understanding the Components
Filter Purpose: Filters detect specific local patterns anywhere in the image.
Bias Purpose: Bias adds a constant offset to influence how easily a neuron activates, essentially adjusting the neuron’s sensitivity threshold.
The Answer
If each filter entry had its own bias, we would be saying that different parts of the pattern have different baseline importance. However, filters are designed to detect complete patterns as unified wholes.
When a filter searches for a pattern (like an edge or corner), all entries work together equally to recognize that pattern. Breaking this unity by assigning different biases to individual entries would compromise the filter’s ability to consistently detect the same pattern across different image locations.
Therefore, we apply one bias value to the entire filter’s output, maintaining the pattern’s integrity while still allowing us to adjust the overall sensitivity of the pattern detector.