Concordia

Teacher-Student neurosymbolic framework, where instead of being a complex deep model, the teacher is a probabilistic logical theory. The framework is implemented in PyTorch. Concordia supports supervised, semi-supervised, and unsupervised training and has been applied to a variety of tasks, outperforming the relevant state-of-the-art. In particular, Concordia outperforms DLP, Bi-LSTM, and DistilBERT on entity linking and IARG and PSL-CAD on collective activity detection when using MobileNet and Inception-v3 as backbone networks. Concordia is strictly more expressive than DLP and T-S in terms of the types of supported logical theories.

Repository

Concordia

Relevant publications