Skip to the lesson

How AI worksLesson 4 of 6

Part 1 of 3, 0 done

14 min

How a computer sees a photo

To a computer, a photo is a grid of numbers. See how a network finds a dent in a bumper, how a change you can't see can fool it, and how image generators paint from static.

Ticket 01 · The guess

A change you can't see

Researchers showed a photo of a panda to an image-recognition network from Google, GoogLeNet. It answered "panda", 57.7%arxiv.org ·2014 sure. Then they changed every pixel by a tiny amount, in a pattern chosen to fool it. To a person, the two photos look the same. The network now said "gibbon", a kind of ape. How sure was it?

50%
0%100%
Words to know

Spot something wrong?

Tell us what looks off. No name or email needed.

What kind of problem?

0 / 1000

We'll also send this page, the lesson, ticket 1, so we can find it. Nothing else. Privacy

·
Sources and every number (1)

Sources

  • Explaining and Harnessing Adversarial Examples (Goodfellow, Shlens and Szegedy; ICLR 2015; figure one, the panda) · Authors, on arXiv, 2014

Every number

  • 57.7%arxiv.org ·2014 · How sure GoogLeNet was that the unchanged photo showed a panda, in the paper's figure one.
  • 99.3%arxiv.org ·2014 · How sure GoogLeNet was that the changed photo showed a gibbon, in the same figure.
  1. Lessons›
  2. The school›
  3. How AI works
  • ·Before this: Training: learning from mistakes
  • ·Next: Why AI makes things up
Skip to the lesson

How AI worksLesson 4 of 6

Part 1 of 3, 0 done

14 min

How a computer sees a photo

To a computer, a photo is a grid of numbers. See how a network finds a dent in a bumper, how a change you can't see can fool it, and how image generators paint from static.

Ticket 01 · The guess

A change you can't see

Researchers showed a photo of a panda to an image-recognition network from Google, GoogLeNet. It answered "panda", 57.7%arxiv.org ·2014 sure. Then they changed every pixel by a tiny amount, in a pattern chosen to fool it. To a person, the two photos look the same. The network now said "gibbon", a kind of ape. How sure was it?

50%
0%100%
Words to know

Spot something wrong?

Tell us what looks off. No name or email needed.

What kind of problem?

0 / 1000

We'll also send this page, the lesson, ticket 1, so we can find it. Nothing else. Privacy

·
Sources and every number (1)

Sources

  • Explaining and Harnessing Adversarial Examples (Goodfellow, Shlens and Szegedy; ICLR 2015; figure one, the panda) · Authors, on arXiv, 2014

Every number

  • 57.7%arxiv.org ·2014 · How sure GoogLeNet was that the unchanged photo showed a panda, in the paper's figure one.
  • 99.3%arxiv.org ·2014 · How sure GoogLeNet was that the changed photo showed a gibbon, in the same figure.
  1. Lessons›
  2. The school›
  3. How AI works
  • ·Before this: Training: learning from mistakes
  • ·Next: Why AI makes things up