CoOL: A Neural Clustering Method Without Distance Computations
Clustering without distance metrics? New paper uses neural agreement instead.
A new paper from Finn Henry O'Shea and Maria Elena Monzani presents Cohort Organized Learning (CoOL), a clustering technique that fundamentally differs from traditional methods by eliminating the need for explicit distance or similarity computations. Instead, CoOL leverages neural networks to learn cluster assignments through an agreement-based mechanism, using expectation maximization to derive gradients for training. The authors provide a thorough derivation of the gradient updates, methods for monitoring convergence during training, and evaluation metrics for post-training cluster quality. They illustrate CoOL's versatility with examples on vector data and image datasets, showing that the method can handle any data modality as long as it can be fed into a neural network.
While promising, CoOL has limitations: the authors note challenges with high-dimensional data and potential sensitivity to initialization. They also discuss future directions, including extensions to hierarchical clustering and semi-supervised settings. The paper, 20 pages with 14 figures and 4 tables, is available on arXiv (2606.21743) and appears in a related DOI in the journal Machine Learning: Science and Technology. CoOL represents a significant shift away from distance-based clustering toward neural agreement frameworks, which could inspire new approaches in unsupervised learning and multiagent systems.
- CoOL uses neural networks to cluster data without explicit distance or similarity calculations.
- Training is driven by expectation maximization, with derived gradients for parameter updates.
- Demonstrated on vector data and images, showing applicability to any compatible data type.
Why It Matters
CoOL offers a fresh neural approach to clustering, potentially enabling more flexible and scalable unsupervised learning.