Dynamic Vector Networks: Self-Organizing Knowledge Structures Beyond Transformers

We introduce Dynamic Vector Networks (DVN), an architecture fundamentally different from Transformers: each node is a rich embedding vector (not a scalar neuron), connections form through Hebbian learning in real-time, and new nodes are spawned dynamically for novel concepts. The architecture combines four properties no existing system offers simultaneously: rich vector nodes, real-time weight evolution, dynamic node creation, and self-organization by semantic similarity.

Also available in French: Dynamic Vector Networks : Au-delà des Transformers