A computer scientist has developed a way to vastly improve the ability of AI to understand images.
Computer scientist Geoffrey Hinton was instrumental in developing image-recognition software used today for everything from transcribing speech to fighting online trolls. In two recent research papers, Hinton has developed a new approach to image recognition software that could help AI systems act more like humans. Hinton’s system aims to remedy a weakness of exiting machine-learning systems, which is that these systems need a large number of example photos in order to reliably recognize even simple objects. This is because the software is limited in its ability to generalise what it learns from one image to another, such as understanding that an object is the same when seen from a new viewpoint. For example, teaching a computer to recognize a dog from any angle requires inputting photos of the dog from thousands of different angles.
Hinton’s idea for fixing this is to create capsules, small groups of crude virtual neurons, which are designed to track different parts of an object, such as the dog’s nose, ears and tail, and their relative positions in space. The capsules can then be networked together to understand and recognise the dog from different views. One of Hinton’s recent research papers demonstrated that his capsule networks can match the accuracy of the best existing software techniques on learning to recognize handwritten numbers. The second paper demonstrated that capsule networks can significantly improve on recognizing toys such as trucks and cars from different angles.
Although Hinton has demonstrated that capsule networks represent a significant improvement in image recognition, the technique remains to be proven on large image collections, and the speed of the system needs to be improved to make it faster than existing software. There have been a wide number of recent developments in AI, such as using AI to detect pre-cancerous growths and to provide detailed marketing analytics. Now, capsule networks may demonstrate a way to make AI operate more like the human brain. What new innovations might arise from AI that can work like the human brain?