# Superposition **Domain:** Machine Learning / Interpretability **Doc Type:** Technical Concept Node **Maturity:** Active Research Concept **Related:** [[Latent Space]], [[Sparse Autoencoder]], [[Monosemantic Feature]], [[Distributed Relational Compression]] --- ## Definition In neural-network interpretability, **superposition** describes a model representing more features than it has directly separable dimensions by encoding features in overlapping directions. An activation direction may therefore participate in several concepts until a suitable decomposition recovers a more interpretable basis. ## Evidentiary boundary This usage is distinct from quantum superposition. It is a representation-learning hypothesis and empirical framework for understanding distributed features in neural networks.