Colloquium Computer Science - Prof. T. Martinetz

Title:
Mulilayer Relation Networks for Abstract Reasoning

Abstract:
Relation Networks as introduced by Santoro et a. in 2017 enable neural networks to cope with logical relations between objects. Relational knowledge, which is seen as the foundation for higher cognition, can now be generated from raw data with end-to-end training combined with CNNs and LSTMs. We give an introduction to these relational networks and give some examples where these networks solve already quite impressive tasks on different benchmark data sets. however, these networks can not cope with higher order relations. To overcome this limitation we introduced the Multilayer Relation Networks, which not only solve these benchmarks perfectly, but also excel on a human IQ test.