Machine Learning for Perception

点击翻译按钮获取中文摘要

Computers are still much worse than people at recognizing objects in images or words in sound-waves. New methods for training artificial neural networks have improved their abilities to learn multiple layers of features and this has led to significant improvements in recognition, but artificial neural networks still lag far behind the real thing. The representational power of neural networks can be greatly improved by grouping neurons into "capsules" that each perform a lot of internal computation and output a small set of real numbers that are bound together as properties of the same entity. By using pairs of images related by a known transformation it is possible to allow each capsule to decide for itself what kind of visual or acoustic entity it should represent but to force each capsule to output the same kinds of coordinates as are used to handle viewpoint effects in computer graphics. This allows spatial relationships to be modeled as linear operations which makes it easy to recognize objects or words by using the relationships between their parts. Low level capsules learn to represent simple entities in small regions of the image. Higher-level capsules learn to represent more complex entities and they operate over larger regions of the image but they retain precise information about the position, orientation and scale of the visual entity they represent. This allows spatial relationships between parts to be used to recognize entities at many different levels in the whole-part hierarchy.