Joint Representation Learning and Clustering via Gradient-Based Manifold Optimization

The Thesis

Clustering high-dimensional data is notoriously hard because distance metrics lose meaning as dimensions grow — a problem known as the curse of dimensionality. Most existing approaches either reduce dimensions first and then cluster, or cluster first and then reduce, meaning each step is blind to the other. This paper proposes doing both at once: a framework that simultaneously tunes a dimension-reduction transform (a linear projection or a neural network) and fits a Gaussian Mixture Model — a probabilistic way of describing data as a blend of several bell-curve-shaped clusters — by walking along a mathematical surface called a manifold. The catch is that the experiments are limited to synthetic data and MNIST, a beginner-level image benchmark that the field largely graduated from a decade ago, so the generalization story is unproven.

Catalyst

Riemannian optimization — the branch of math that lets gradient descent work on curved surfaces rather than flat Euclidean space — has matured into practical software libraries (Pymanopt, Geoopt) over the past few years, lowering the implementation barrier for manifold-constrained learning. Simultaneously, self-supervised and unsupervised representation learning have become central research topics, creating renewed demand for principled joint-learning frameworks. These two currents converging made this specific construction tractable now in a way that would have required much heavier custom engineering five years ago.

What's New

Earlier approaches to joint clustering and dimensionality reduction — such as Deep Clustering (which alternates between updating a neural network and reassigning cluster labels) and Spectral Clustering (which embeds data via graph eigenvectors before clustering) — treat the two objectives sequentially or in loose alternation, meaning gradients from the clustering objective rarely reach the representation parameters cleanly. This paper proposes a single unified loss surface and uses Riemannian gradient descent — gradient steps that stay on the manifold of valid projection matrices — to optimize both objectives simultaneously. The authors claim this avoids the instability and information loss that alternating updates introduce.

The Counter

The core idea — joint optimization of representation and clustering — is not new. Deep Clustering (Caron et al., 2018), DEC (Xie et al., 2016), and several follow-ons have explored this space with much larger-scale experiments and more honest baselines. Beating k-means and vanilla GMM on MNIST in 2024 is not a meaningful result; those baselines were surpassed years ago. The Riemannian optimization machinery adds theoretical elegance but also real computational cost, and the paper provides no evidence that this cost is justified by downstream performance on anything harder than a toy dataset. The Gaussian Mixture Model assumption is baked into the entire framework, which means the method will struggle anywhere data clusters are non-Gaussian — which is most places that matter. Without open code, ablations on real-world datasets, and comparisons to modern deep clustering methods, the paper's claims rest on a very narrow experimental foundation.

The Paper