Data Science Analogies

Data Science Analogies: The Blind Men and the Elephant

Convolutional neural networks for image recognition remind me of the parable of the blind men and the elephant. The higher layers of the network will try to detect simple features based on the edges and textures detected by deeper layers. Then, dense layers at the top of the network will try to make a prediction of what the image is based on the features found by the layers below it. In the case of the elephant parable, the “features” found were a rope, a tree, and a snake – what do you think this human CNN would have predicted as the whole picture? A rope swing hanging from a snake-infested tree?

An alternative scenario – what if each person said, “This could be a really big tail, so there should be two really big legs somewhere that way – and if not, then it’s probably a rope.” That’s like Geoffrey Hinton’s concept of “capsule networks”. Each feature contains a sense of where other related features should be given its “pose”.

Picture from The Heath readers by grades, D.C. Heath and Company (Boston), p. 69. Text from Wikipedia: https://en.wikipedia.org/wiki/Blind_men_and_an_elephant Hinton explaining capsules: https://www.youtube.com/watch?v=6S1_WqE55UQ