Geoffrey Hinton, a Google researcher and professor at the University of Toronto, helped pioneer artificial neural networks, the technology behind most of the major advances in machine learning. And now he's come up with a new idea that he thinks is even more powerful.
Hinton calls his latest creation "capsule networks." Each capsule is a group of artificial neurons trained to track a specific feature of an image. Combining them allows an AI system to understand the spatial relations between different features of an image, so it can identify different views of the same image. Hinton has shown that this technique performs much better than existing systems in a challenge to recognize objects from different angles.
Why it matters: To existing neural networks, two images of the same object from different angles look like totally different objects. This means neural networks asked to recognize objects in images need to train on images from many different angles, which requires vast amounts of data. For example, the ImageNet data set, used in the image recognition competition that's been the benchmark for these systems for the last seven years, contains more than 13 million images. The hope is that capsule networks could achieve the same results working from much smaller data sets.