siddk/entity-network

Implements a multi-cell memory architecture with entity-specific tracking, where each memory cell maintains both a key (entity identifier) and hidden state (content) updated via gated mechanisms similar to GRUs. Built with TensorFlow/TFLearn, the model combines a learned multiplicative input encoder, dynamic memory module with key-value separation, and an attention-based output decoder that weights memories by cosine similarity to queries. Evaluated on bAbI question-answering tasks, demonstrating interpretable world-state maintenance by allowing fixed entity embeddings as memory keys.

No commits in the last 6 months.