Brain-inspired 'configurations' model achieves 87% AUC in novelty detection

In a new preprint, researchers from multiple institutions introduce 'configurations,' a clustering framework inspired by how infants discover categories and detect novelty without supervision. Unlike traditional clustering methods that require specifying the number of clusters, configurations uses a single resolution parameter and attraction-repulsion dynamics to produce hierarchical groupings, adapt to new contexts, and flag outliers. The authors also developed mheatmap, a tool for proportional heatmaps and reassignment algorithms that fairly evaluates multi-resolution and dynamic behavior.

Across various datasets, configurations matches or beats standard clustering metrics, demonstrates strong novelty detection with 87% AUC, and shows 35% better stability when categories evolve over time. The work positions configurations as a principled computational model of early cognitive categorization, offering a step toward brain-inspired AI that can learn and adapt like humans.