A bank reviews photos of paper checks to flag a small altered-ink pattern that can sit anywhere on the check. The same local mark should be recognized in the corner or in the middle. Which network design matches that job?
Select an answer to reveal the explanation.
Short Explanation
The same ink smudge can land in any corner of a check, like a stencil you slide across the page. You want a convolutional network that reuses one local filter at every spot on that image grid. The trap is a fully connected net that learns a separate weight for each pixel position.
Full Explanation
The altered ink is a local pattern on a grid, and its meaning does not depend on which corner of the check holds it. A convolutional layer slides a small filter across that grid and reuses the same weights at every location, so a mark learned in one place is recognized in another. Shared weights are why a mark in the corner and the same mark in the middle are one problem rather than two unrelated pixels. That shared local filter is the mechanism that matches the job. A fully connected layer assigns a separate weight to each pixel position, so a pattern in the middle and the same pattern in the corner are unrelated inputs. A recurrent scan updates one hidden state pixel by pixel and does not reuse a two-dimensional local filter across the image. Clustering on average brightness collapses the check to one global number and cannot hunt for a small repeated mark. Exam caveat: a dense net can memorize a mark that always sits in one fixed spot; it does not share a filter when the mark can move. Operational check: confirm the first learned layer is a sliding kernel with weights tied across locations, not a dense map from every pixel to the next layer.