Speaker: Alex Krizhevsky
Department of Computer Science
University of Toronto
Title: Image Retrieval using Short Binary Codes found by Deep Learning
Abstract: We use DBNs to learn many layers of features on color images, and then
use these features to initialize a very deep autoencoder. We then use
this autoencoder to learn a mapping between images and binary codes.
The binary codes allow very fast retrieval of images that are similar
to a query image. These codes tend to capture something of the global
appearance of the image, but they don't capture what might be called
the "human similarity metric". For example, humans would consider
similar two images of a cat and a person in reversed positions. We
show that we can go some way towards capturing this metric by training
a model on pieces of an image.
For Additional Information, Contact: