When would you replace every 2x2 max pooling layer with a stride-2 convolution, and what does it cost?

## The two ways to halve a feature map **Fixed reduction.** A 2x2 max pool with stride 2 takes each disjoint 2x2 window of each channel and emits its maximum. No parameters, negligible compute, and a hard-coded assertion about what matters: the strongest response in the neighbourhood. **Learned reduction.** A convolution with stride 2 — commonly 3x3 with padding 1, which maps an even n to exactly n/2 — slides its kernel two positions at a time. It produces the same spatial halving, but each output is a trained weighted sum over the window and across all input channels, followed by the usual nonlinearity. The network learns what summarising means here. An architecture built this way, with no pooling layers at all and every reduction performed by a strided convolution, is what "all-convolutional" refers to. ## What the swap buys **Expressiveness.** Max is one statistic. A learned kernel can approximate a blur, an edge-preserving reduction, a channel-selective summary, or something with no name — and it can learn a different rule at every stage and every channel group. Where the useful reduction is not "take the peak", pooling cannot represent it and the strided convolution can. **Fusion.** The reduction is not a separate layer. You needed a convolution at that point in the block anyway; giving it a stride of 2 removes a layer from the graph and performs the downsample as part of work already being done. It also means the downsample sees all input channels jointly, whereas pooling is strictly per channel and can never mix them. **Control of capacity at the reduction.** Because the output channel count is yours to choose, the standard move is to double channels at each spatial halving, keeping the per-position information budget roughly constant as the grid shrinks. Pooling cannot change channel count at all, so that adjustment needs another layer. ## What the swap costs **Parameters.** A stride-2 3x3 convolution from 64 to 128 channels holds 3 * 3 * 64 * 128 = 73,728 weights plus 128 biases. Pooling holds zero. Repeat that at four or five reduction stages and the difference is real, especially on a device budget. **Compute.** Multiply-adds equal the parameter count times the number of output positions. Striding halves each spatial dimension, so the reduction runs on a quarter of the positions of a stride-1 layer, but it is still vastly more work than a comparison per window. **A lost prior.** Fixed max pooling encodes a genuinely reasonable assumption at zero cost. Removing it means the network must learn the reduction from data; with a small training set that is capacity spent relearning something you already knew, and it can overfit. The swap is most defensible when data is plentiful relative to the model. **No free shift stability.** Neither operator escapes aliasing. Both sample on a fixed grid: shift the input by one pixel and the set of positions retained changes, so downstream features can change more than the shift warrants. Anti-aliased downsampling — low-pass filtering before subsampling, in the spirit of blur-pooling — attacks that directly and is orthogonal to the pooling-versus-strided-convolution choice. A candidate who claims strided convolutions fix aliasing has the wrong model of the problem. ## How to decide Ask three questions. 1. **How much data do you have relative to model size?** Plentiful data favours the learned reduction; scarce data favours keeping the free prior. 2. **What is the budget?** On a tight parameter or latency budget, pooling is a genuinely attractive zero-cost reduction, and the compute you save can go into a wider trunk instead. 3. **Is the max the right summary for this signal?** If the class evidence is peaky, the fixed max is already doing the right thing and a learned kernel may only rediscover it. If the evidence is distributed or the useful summary is channel-dependent, the learned reduction can express something pooling cannot. And measure. The difference between the two is typically modest and architecture-dependent, so the honest senior answer is that this is a change you A/B under a fixed parameter or latency budget, not a rule you apply everywhere. Swapping in strided convolutions while also adding parameters and then attributing the gain to the swap is the classic confounded experiment; hold the budget constant or the comparison says nothing. ## One place the choice is nearly free The reduction immediately before a global pooling head matters least, because whatever survives is about to be averaged over the whole map anyway. The reductions that matter are the early ones, where a great deal of spatial information is being discarded and the rule that decides what to keep still has a lot of signal to choose from.