Graph Neural Networks (GNN) in logiCells

Documentation status: reference — see Maturity and evidence.

Graph neural networks learn from node features together with graph structure.

logiCells specificity

The source structure can come from the conceptual graph/hypergraph, allowing learned representations to remain connected to explicit concepts and relations.

Typical computation

A GNN layer usually combines:

1. node features;
2. neighborhood structure;
3. an aggregation rule;
4. a learnable transformation;
5. an activation function.

Normalization based on node degree or another graph operator can stabilize propagation.

Conceptual adjacency

Adjacency or incidence structures can be generated from conceptual relations, then projected into numerical forms suitable for the selected neural kernel.

Learning

GNN parameters are trained through the same differentiable computation principles as other neural layers. The conceptual graph remains the semantic source; learned tensors are execution projections.