Even from a blurry photo

Blurry, but still recognized

Key ideaEven a blurry clue finds the closest memory

This time we show a blurry photo of the house, with 40% of its dots flipped at random. Only 61% of the dots match the original. Still, the network fixes the wrong dots one by one along its links and brings the house back exactly. That is why you still recognize a day in an old, faded photo.

Why?The network slides from the clue toward the closest picture it has stored. It is like a ball on a hillside rolling down into the nearest valley. So even if the clue is a bit wrong or a bit incomplete, it reaches the same memory.

Go furtherBecause each stored picture becomes a valley, networks like this are also called attractor networks. If a clue is too blurry, it can roll into a different valley and bring back the wrong memory.