HomRMN

HomRMN, short for Homogeneous Relational Memory Network, is a neural network architecture designed for memory-augmented reasoning on homogeneous graphs. It operates on graphs where nodes and edges reside in a common feature space and shares parameters across all nodes and relations, enabling scalable, uniform processing of structure with a single memory scheme.

The core architecture consists of a memory bank divided into slots, typically one slot per node in the graph. Each slot stores a latent representation that is updated over a series of computational steps. A read mechanism uses attention over the memory slots to form a contextual vector, from which a control unit determines how to update the memory. Updates are performed through gated mechanisms that combine prior memory with new message-passed information from neighboring nodes, allowing multi-hop reasoning without separate per-relations parameters. The process unfolds recurrently, with the same memory and update rules applied at each step, facilitating iterative refinement of node representations.

Training and objective functions for HomRMN are task-dependent. Common goals include node classification, link prediction, or graph-level property estimation, with losses such as cross-entropy or regression error. Training typically uses backpropagation through time or unrolled computation over a fixed number of steps, often with regularization to prevent overuse of the memory. The homogeneous design supports weight sharing, which can improve generalization across graphs of varying size.

Applications of HomRMN span molecular and social graphs, knowledge graphs with uniform relation types, and other domains requiring relational reasoning over structured data. Its advantages include a compact memory scheme and the ability to perform iterative reasoning, while limitations involve scalability to very large graphs and potential challenges in capturing highly heterogeneous interactions. Related concepts include Relational Memory Networks and graph neural networks.