STRATA
Skip to study
XLVII

IV.09 / AI

Deep learning and the 2012 hinge

How much of the change in image recognition can be attributed to training data, accelerators and architecture?

A source-led introduction1 source references0 related experiments

Start with the question.

The 2012 ImageNet paper connected a deep convolutional architecture with a large supervised image benchmark and a practical training system. Its historical significance is clearest when architecture, data, optimization, and computational resources are examined together. This topic asks which changes actually improved a measured task and which lessons survive when the images or deployment conditions change.

Inspect a classifier's failures by category.

Create a small evaluation set with ordinary images, unusual viewpoints, and altered backgrounds. Keep the labels and source records visible. Compare errors across these conditions instead of reporting only one accuracy value; a model that relies heavily on background cues may succeed on the familiar set and fail when the object remains unchanged.

What would you inspect?

  • Training and evaluation distributions, preprocessing, and label provenance.
  • Per-class errors, confidence calibration, and performance under controlled image changes.

Take the next step.

Reconstruct one component comparison from the original system description and identify the additional experiment needed to isolate its contribution.

Read the supporting material.

  1. ImageNet Classification with Deep Convolutional Neural Networks

    Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton / NeurIPS · The 2012 convolutional-network system and ImageNet evaluation.

Follow the comparisons.

Read the neighboring questions.