AlexNet: the same recipe at ImageNet scale

AlexNet (2012) is the digit reader scaled up to a thousand kinds of photograph; here a pre-trained image network runs live in the browser on test pictures or your camera.

This is the text of a chapter of the animated talk A Deep Seek into LLM Architecture by Ruben Galvão. Open the page with scripts on to play the slides, one key press per idea.

AlexNet (2012)

• 2012, built for ImageNet: 1.2 M photos, 1,000 classes
• Same recipe: 227×227×3 in; conv, ReLU, pool; channels 96, 256, 384, 384, 256
• Flatten 9,216 → 4,096 → 4,096 → 1,000 + softmax
• ≈ 60 M weights (10,000× the digit reader); ≈ 1.4 G ops per image (≈ 3,650×)
• Maps after the first are illustrative

The real weights

• The real AlexNet (61 M weights) running in this page, offline
• Centre crop only (white frame); top-1 ≈ 55–57 %, so we show top five
• Hosted version runs MobileNetV2 (≈ 3.5 M) instead
• 2012: top-5 error 15.3 % vs 26.2 % for the runner-up: deep learning takes over vision
• Next: if depth helped, add more

Papers and sources

• Deng et al. (2009): ImageNet: A large-scale hierarchical image database
• Krizhevsky, Sutskever & Hinton (2012): ImageNet Classification with Deep Convolutional Neural Networks (AlexNet)
• BVLC · ONNX Model Zoo (pre-trained weights): The Caffe reference AlexNet, converted to ONNX