👟 AlexNet Teaching Lab

Shapes • Params • FLOPs • Receptive Field

Step 1 — Choose input & presets

Load an image (scaled to AlexNet size), then explore layers and math.

227×227

What is AlexNet?

Paper
Krizhevsky, Sutskever, Hinton (2012)
Key ideas
Large kernels early (11×11 s4), ReLU, LRN, MaxPool, deep conv stack, big FC layers, Dropout
Input
~227×227×3 image
Output
1000-way ImageNet classification
ReLU nonlinearity Local Response Norm (LRN) MaxPool 3×3 s2 Dropout in FC

Step 2 — Shapes, Params, FLOPs, Receptive Field

These are computed live from the classic AlexNet hyperparameters.

LayerKernel / Stride / PadIn (H×W×C)Out (H×W×C)ParamsFLOPs (≈)RF

Step 3 — Toy forward pass (Conv1 → ReLU → Pool1)

We apply a small set of handcrafted Conv1-like filters to illustrate feature maps (for speed).

Input
Conv1 (toy)
After Pool1

Note: Real AlexNet uses 96 learned 11×11×3 filters at stride 4 (we show a tiny subset on a downscaled buffer).

Try changing the input size to see how every layer’s spatial size, parameter count, FLOPs, and receptive field change. The toy forward pass demonstrates how early filters respond to edges and textures, and how pooling reduces resolution while gaining invariance.