Self-Organizing Maps (SOMs) excel at providing topology-preserving projections of high-dimensional data, yet their capacity for interpretable latent navigation and generative modeling remains largely unexplored. In this work, we show that the activation pattern of a SOM - the squared distances to its prototypes — can be exactly inverted to reconstruct the original input space. This follows from a classical result in Euclidean distance geometry: a point in $D$ dimensions is uniquely determined by its distances to $D+1$ affinely independent references. Leveraging this geometric foundation, we introduce MUSIC (Manifold-Aware Unified SOM Inversion and Control), an update rule that enables interpretable navigation across the learned manifold. By modifying squared distances to target prototypes while preserving those to other prototypes, MUSIC generates controlled, semantically meaningful trajectories aligned with the SOM's piecewise-linear structure. Tikhonov regularisation is introduced to stabilize the inversion and ensure smooth, continuous motion in high-dimensional spaces. Unlike variational or diffusion-based generative models, our framework requires no stochastic sampling, latent priors, or learned decoders; it operates entirely on the map's intrinsic geometry. When traversing between distinct clusters, MUSIC produces coherent semantic transitions rather than naive pixel blending. We validate our framework on synthetic Gaussian mixtures, MNIST digits, and the Labeled Faces in the Wild dataset. Across all settings, these interpretable promenades maintain high classifier confidence, yield significantly sharper intermediate images than linear interpolation, and expose the rich geometric structure of the underlying map.
Navigating Self-Organizing Maps: Interpretable Generation via Geometric Inversion / Londei, A., Benati, M., Lanzieri, D., Loreto, V.. - (2026). [10.2139/ssrn.7118514]
Navigating Self-Organizing Maps: Interpretable Generation via Geometric Inversion
Alessandro Londei;Matteo Benati;Vittorio Loreto
2026
Abstract
Self-Organizing Maps (SOMs) excel at providing topology-preserving projections of high-dimensional data, yet their capacity for interpretable latent navigation and generative modeling remains largely unexplored. In this work, we show that the activation pattern of a SOM - the squared distances to its prototypes — can be exactly inverted to reconstruct the original input space. This follows from a classical result in Euclidean distance geometry: a point in $D$ dimensions is uniquely determined by its distances to $D+1$ affinely independent references. Leveraging this geometric foundation, we introduce MUSIC (Manifold-Aware Unified SOM Inversion and Control), an update rule that enables interpretable navigation across the learned manifold. By modifying squared distances to target prototypes while preserving those to other prototypes, MUSIC generates controlled, semantically meaningful trajectories aligned with the SOM's piecewise-linear structure. Tikhonov regularisation is introduced to stabilize the inversion and ensure smooth, continuous motion in high-dimensional spaces. Unlike variational or diffusion-based generative models, our framework requires no stochastic sampling, latent priors, or learned decoders; it operates entirely on the map's intrinsic geometry. When traversing between distinct clusters, MUSIC produces coherent semantic transitions rather than naive pixel blending. We validate our framework on synthetic Gaussian mixtures, MNIST digits, and the Labeled Faces in the Wild dataset. Across all settings, these interpretable promenades maintain high classifier confidence, yield significantly sharper intermediate images than linear interpolation, and expose the rich geometric structure of the underlying map.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.


