In a CNN, what is the real cost of choosing valid padding over same zero padding?

## The two modes **Valid padding** means no padding at all. The kernel is placed only where its entire window sits inside real input, so the output is smaller than the input: a 3x3 kernel at stride 1 loses one pixel from each of the four sides, a 5x5 kernel loses two. **Same zero padding** rings the input with a border of zeros - one ring for 3x3, two for 5x5 - so that a stride-1 layer emits a map with the same height and width as its input. "Same" is a statement about stride-1 behaviour; with a stride above 1 the output shrinks by the stride regardless of padding, which is a routine misunderstanding. Neither mode introduces learnable parameters. Padding is data manipulation, not capacity. ## What valid padding costs: a shrinking, blind border The shrinkage compounds. Ten stacked valid 3x3 layers remove ten pixels from every side. Beyond the bookkeeping annoyance, the substantive cost is that the outer ring of the input has no output position that is centred on it. Content there still influences some outputs - it appears in windows centred one pixel inward - but it can never be the subject of a prediction. On a scanned document field this is the failure that bites: the signature's underline sits in the bottom few pixel rows of the crop, and after a valid stack there is simply no output unit whose job is that region. The model behaves as though the field ends slightly higher than it does. Anything where evidence hugs the frame - marks at the edge of a form, structures at the boundary of a tile - is exposed to the same problem. The shrinkage also makes architecture with skip connections awkward: two paths that started the same size no longer match, so a valid-padded design has to crop one side to align them. ## What same padding costs: fabricated context Same padding buys size preservation by inventing values. A window centred on a corner pixel of a 3x3 layer contains five real values and four zeros. Its output is therefore a blend of measurement and fiction, and border statistics differ systematically from interior statistics. There is a subtler consequence. The zero border is a constant, reliable signal that appears only near the edges, and a convolutional stack can learn to use it to infer *absolute* position - how far a unit is from the frame - even though convolution is nominally position-agnostic. Sometimes that is useful. Often it is an invisible shortcut: a model trained on centred, uniformly framed crops can lean on the border cue and degrade when inputs are tiled, cropped differently, or presented at another size. Zeros are also not the only option. Reflecting or replicating the edge values keeps the border statistics closer to the real distribution and avoids injecting a hard discontinuity, at the cost of duplicating information the model may then over-count. ## Choosing Use **same** padding when geometry must be preserved: dense prediction where an output pixel should correspond to an input pixel, architectures with additive skip connections, or deep stacks where valid shrinkage would eat the image. Use **valid** padding when fabricated values are genuinely unacceptable. The classic setting is tiled inference over an image far too large to process at once: you run the network on overlapping tiles and stitch the results, and you want every output to depend only on real data so that seams do not show. A design in that spirit uses unpadded convolutions throughout and supplies the missing context by overlapping the tiles rather than by inventing it - and at the outer boundary of the whole image, where no real context exists, it mirrors the image rather than zeroing it. ## Interview framing The weak answer is "same keeps the size, valid shrinks it" and stops there. That is the definition, not the cost. The strong answer names both prices: valid gives up a border region that ends up unrepresented in the output, and same gives up border authenticity and hands the model a positional shortcut it did not ask for. The strong answer also flags the trap that same padding does not preserve size once the stride is above 1. ## Diagnostic tells If a segmentation model is systematically worse in a band along the image edge, look at padding first. If a model trained on full frames collapses on cropped or tiled inputs, suspect it learned something from the constant border. And if two branches of a network refuse to add, count the valid convolutions on each path.