4 ms·

What is meant by 'routing'?

• JellyBeanThief 4y ago
  > A longstanding goal in Artificial Intelligence is to formulate learning systems that assign credit, such that, when they succeed or fail in a task, we can determine and interpret which components of the system are responsible. A possible approach to the credit assignment problem is to route capsules at multiple levels of composition. A capsule is a group (e.g., vector, matrix) of artificial neurons representing the properties of an entity in a context (e.g., a token of text in a paragraph, an object depicted in an image). Routing consists of assigning data from input capsules, each representing a detected entity, to output capsules, each representing a detectable entity, by finding or computing agreement in some form (e.g., identifying clusters) among candidate output capsules proposed by transforming the input capsules. Each output capsule is computed as a mixture of the candidates proposed for it on which the most input capsules agree, thereby assigning credit to those input capsules. If we compose multiple routings into a deep neural network, in every forward pass it assigns credit to input capsules representing the entities detected at each level of composition. I hope that helps you, because it didn't help me. I think it's about trying to expose how a neural network makes decisions by saying these neurons over here were responsible for making those neurons over there be the way that they are. I really need to get into AI.