liveBook · Manning

Robustness

As explained in Sutskever's List

Robustness is a model's ability to maintain reliable behavior under perturbations, distribution changes, and atypical inputs.

A model can surpass human-level ImageNet scores yet fail on inputs that humans recognize immediately. Robustness asks whether reliable behavior survives perturbations, distribution shifts, and atypical examples. section 3.6

CNNs exposed this gap through adversarial examples: imperceptible changes could trigger confident misclassifications. Similar failures appeared in deployment, where underrepresented cases and subtle real-world shifts defeated systems that performed well on benchmarks. section 3.6 section 3.6

Robustness failures include:

  • Treating benchmark accuracy as evidence of general visual understanding. section 3.6
  • Misclassifying imperceptibly perturbed inputs with high confidence. section 3.6
  • Producing harmful errors on underrepresented examples, such as Google Photos labeling Black individuals as “gorillas.” section 3.6

How Sutskever's List covers it

How deep it goes
Identifies adversarial and deployment failures but does not develop robustness mechanisms or detailed evaluation procedures.
Examples
Uses ImageNet scores, imperceptible perturbations, and Google Photos’ gorilla-labeling incident.
What it assumes
Assumes familiarity with CNNs, ImageNet benchmarks, and classification errors.
What you can do after
Evaluate claims of human-level performance against adversarial inputs and atypical real-world cases.
How it connects
Connects robustness with adversarial examples, generalization, distribution shift, interpretability, and convolutional neural networks.
Cover of Sutskever's List
liveBookmanning.com
Read Sutskever's List on liveBook Get the book at manning.com

Also explained in

Compare how 7 books explain robustness →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Sutskever's List. The section links point to the parts of the book it draws on.