Hello everybody,
today (3 March 2020) at 6 pm, Johannes Rabold will give a talk on "Capsule Networks" in the CogSys Lab (WE5/05.013):
Modern Convolutional Neural Networks are around since the 80s and since then completely changed the game of Computer Vision. The talk will first shortly revise the general ideas of that architecture and show critical shortcomings of it from the perspective of human visual perception. Capsule Networks proposed by Geoffrey Hinton try to get rid of these shortcomings by introducing the idea of capsules into the system: Capsules do not encode feature detectors for every variant of an entity but rather the probability of presence as well as a high dimensional pose information. Capsule Networks have been applied to standard datasets like MNIST or NORB and show state of the art results. Their architecture makes it also possible to gain insight into the network routing of information and thus making them more interpretable than CNNs.
You are all very welcome to join.
Best regards
Mark Gromowski