Computers can drive cars, beat the champions of Chess and World Chess, and even write articles. Much of today’s AI revolution is due to technological advances known as convolutional neural networks, or CNN. CNN specializes in learning and identifying patterns in two-dimensional plane data.
But CNN doesn’t work well, but if the dataset isn’t based on flat geometry, but is something like an irregular model used in 3D animation, or a point cloud that is generated when a self-driving car draws its surroundings. Around 2016, a new discipline called Geometric Deep Learning is trying to help CNN get out of the plane.
Now, the researchers have come up with a new theoretical framework that allows AI to see a higher dimension. Known as the specvoluldonic neural network (gauge-equivariant convolutional network or gauge CNN), it detects not only 2D pixel arrays, but also patterns in spheres and asymmetric surface objects. The norm CNN has a deep connection to physics.