Neural Operators Can Discover Functional Clusters

A research team led by Yicen Li has published a groundbreaking paper proving neural operators can perform universal clustering on infinite-dimensional functional data, fundamentally expanding the capabilities of operator learning beyond regression tasks. Their work establishes that sample-based neural operators can learn any finite collection of classes in reproducing kernel Hilbert spaces, even when those classes are neither convex nor connected, under mild kernel sampling assumptions. This represents a significant theoretical advance, as clustering in infinite-dimensional spaces has remained poorly understood compared to regression applications of neural operators.

The team's universal clustering theorem demonstrates that any K closed classes can be approximated to arbitrary precision by NO-parameterized classes in the upper Kuratowski topology on closed sets, which mathematically guarantees zero false-positive misclassifications. Building on this theoretical foundation, they developed a practical neural operator-powered clustering pipeline for functional data and applied it to unlabeled families of ordinary differential equation trajectories. Their implementation, called SNO (Sample-based Neural Operator), lifts discretized trajectories via a pre-trained encoder into continuous feature maps, then maps them to soft assignments through a lightweight trainable head. Experiments on diverse synthetic ODE benchmarks show SNO successfully recovers latent dynamical structure in regimes where classical clustering methods fail, providing empirical validation of their theoretical results.