Share one filter
A convolutional layer applies the same filter at multiple positions. Its output is a feature map. Sharing weights makes the same local pattern detectable in different locations, though padding, stride and later layers affect the exact behavior.
Build a larger view
A network can combine feature maps across layers. Early filters may respond to simple local patterns; later layers can combine information from larger regions. These are useful intuitions, not a guarantee that every filter learns a named feature. Change the second input and follow the shared filter across adjacent windows. ReLU then replaces negative filter responses with zero.
INTERACTIVE EXPLORATION
Scan and activate
Change the second input in [1, x, 2, 4]. Apply [1, −1] to each valid window, followed by ReLU.
Try the detector
Our experiment slides a fixed [1, −1] filter over four numbers and then applies ReLU. The filter detects local differences; ReLU keeps positive responses. It is one tiny step, not a complete trained image classifier.
Check your understanding
What does ReLU do to a filter output of −2?
It returns 0, because ReLU(z) = max(0, z).