What does AlexNet's 11x11 stride-4 first layer discard that a 3x3 stem keeps?

## Two knobs, often conflated AlexNet's first layer uses 11x11 kernels at stride 4; VGG's first layer uses 3x3 kernels at stride 1. Candidates usually answer as if that is one decision, but kernel size and stride do different damage and buy different things, and a good answer separates them. ## What the stride costs Stride 4 means the layer evaluates its filter only at every fourth position, so the output map is a quarter of the input's size in each dimension right after layer one. The sampling grid is now coarse, and anything that varies faster than that grid can resolve is gone — folded, in the aliasing sense, into whatever the filter happened to average over. That loss is permanent: later layers see only the subsampled map, and no amount of depth reconstructs structure that was never measured. Tasks that depend on fine detail feel it first — small objects, thin structures, fine texture, precise localization. For coarse whole-image classification the loss is often affordable, which is precisely why the design survived as long as it did. ## What the kernel size costs An 11x11 filter over 3 input channels computes one weighted sum over a wide window and passes it through one nonlinearity. Whatever pattern it responds to must be expressible as a single linear template plus a pointwise nonlinearity. Stacked small kernels reach comparable window size through several linear-plus-nonlinear stages, which is a strictly richer composition per unit of reach. This is the same factorization argument that makes two 3x3 layers preferable to one 5x5 layer, applied at the stem. ## What the aggressive stem buys Compute. A convolution layer costs about `k*k*C_in*C_out*H*W` multiply-adds, and `H*W` is largest at the input. Striding by 4 at the stem shrinks `H*W` by roughly sixteen for that layer and for everything after it. The saving is enormous and it is why an early network could be trained at all on the hardware of its day. VGG's decision to convolve at full resolution for the first block is the opposite bet: those layers hold a trivial number of weights — a 3x3 layer at 64 to 64 channels is under forty thousand — and yet, running on a 224x224 map, they consume a large share of the network's arithmetic. ## Why the stem is not where the parameters are A common wrong answer blames the 11x11 kernels for AlexNet's size. Count it: 11 times 11 times 3 input channels times the filter count is a few tens of thousands of weights. Like VGG, AlexNet keeps the overwhelming majority of its parameters in the fully connected layers at the end. Large kernels are expensive in arithmetic at high resolution, not in storage at the stem. ## The judgment to demonstrate The stem is a budget decision, and framing it that way is what separates a senior answer from a recital. You are choosing how much resolution to spend before the network has learned anything, and the right answer depends on the downstream task and the latency budget rather than on a rule. A coarse, cheap stem is right when the task is whole-image and the compute budget is tight. It is the wrong call when the task needs fine localization, because the first layer sets a ceiling on precision that nothing later can lift. In between, the usual compromise is to reach the same output stride over several layers rather than in one jump, so that some nonlinear processing happens before the detail is thrown away. ## The trap in the factorization argument One last subtlety. Replacing an 11x11 stride-4 layer with three stride-1 3x3 layers does reduce weights, but it multiplies the arithmetic, because those layers now run at full resolution. The cheap-small-kernels argument compares kernels at equal spatial size; it says nothing about downsampling. If you want the stem's compute saving, you have to keep the downsampling — you can only spread it out.