Geoffrey Hinton of Google Brain, considered the father of Deep Learning, has published research articles that have introduced a new type of capsule-based neural network. In addition, he published the algorithm, called “Dynamic Routing Between Capsules”, which allows to have a new alternative to traditional neural networks: Dynamic Routing between Capsules.
These capsule-based neural networks is a breakthrough in the world of deep learning as it allows for the correction of defects in convolutional neural networks that are widely used for image recognition. Indeed, with capsule-based neural networks, these networks have the capacity to take into account contextual information on the characteristics extracted from the image, in particular their spatial and hierarchical organization within the image. For example, if we drive a convolutional neural network to detect faces on images, it will detect a human face on a painting that represents a cubic face of Picasso, and so it is an error and a limitation of convolutional neural networks . On the other hand, with a network of capsule-based neurons that takes into account the spatial and hierarchical organization within an image, it will not detect a human face on this Picasso painting, and that’s what makes this type of neural network more efficient than the convolutional.
An implementation with PyTorch of this algorithm is available on GitHub:
Be the first to comment