Unsupervised Feature Learning with Emergent Data-Driven Prototypicality

Yunhui Guo, Youren Zhang, Yubei Chen, Stella X. YuOpenReview 2023 · Submission to ICLR 2024 · Withdrawn · OpenReview

Abstract

Given an image set without any labels, our goal is to train a model that maps each image to a point in a feature space such that, not only proximity indicates visual similarity, but where it is located directly encodes how prototypical the image is according to the dataset. Our key insight is to perform unsupervised feature learning in hyperbolic instead of Euclidean space, where the distance between points still reflect image similarity, and yet we gain additional capacity for representing prototypicality with the location of the point: The closer it is to the origin, the more prototypical it is. The latter property is simply emergent from optimizing the usual metric learning objective: The image similar to many training instances is best placed at the center of corresponding points in Euclidean space, but closer to the origin in hyperbolic space. We propose an unsupervised feature learning algorithm in Hyperbolic space with sphere pACKing. HACK first generates uniformly packed particles in the Poincar'e ball of hyperbolic space and then assigns each image uniquely to each particle. Images after congealing are regarded more typical of the dataset it belongs to. With our feature mapper simply trained to spread out training instances in hyperbolic space, we observe that images move closer to the origin with congealing, validating our idea of unsupervised prototypicality discovery. We demonstrate that our data-driven prototypicality provides an easy and superior unsupervised instance selection to reduce sample complexity, increase model generalization with atypical instances and robustness with typical ones.

Introduction: The Challenge of Prototypicality

In the world of unsupervised learning, the primary objective is typically to group similar data points together. We want the features of a "cat" to be near other "cat" features in a latent space. However, not all samples are created equal. Some are "prototypical"—the platonic ideal of a category—while others are outliers: blurry, oddly angled, or partially obscured.

Identifying these representative samples (prototypes) is usually a complex task involving heuristics or supervised labels. A new approach, HACK (Hyperbolic space with sphere pACKing), introduces a method where prototypicality becomes an "emergent" property of the underlying mathematics. By shifting the geometry of the feature space from Euclidean to Hyperbolic, the most representative images naturally gravitate toward the center of the representation.

Comparison between traditional unsupervised learning and the HACK approach, showing how HACK organizes data structurally.

Methodology: The HACK Approach

The core innovation of this research is the use of Hyperbolic Space, specifically the Poincaré ball model, as a structural prior for feature mapping.

Why Hyperbolic Geometry?

In standard Euclidean space, the "center" is not inherently special. In a Poincaré ball, however, the distance to the origin grows exponentially as one moves toward the boundary. This geometry is naturally suited for hierarchical structures, where the origin acts as a "root" that is relatively close to all other points. This property allows the model to encode prototypicality through radial distance: the closer a point is to the origin, the more typical it is of the dataset.

Sphere pACKing and Instance Assignment

The algorithm begins by generating a set of "particles" that are uniformly packed within the hyperbolic ball. These particles serve as fixed target locations for the training images.

Uniformly packed particles in the Poincaré ball used as fixed targets for unsupervised feature learning.

During training, HACK employs the Hungarian algorithm—a global bipartite matching technique—to assign each image in a batch to a unique particle. The optimization objective is to minimize the total hyperbolic distance between images and their assigned targets. Because the origin is the most "accessible" point in hyperbolic space, the optimization naturally assigns the most "central" or representative images—those sharing features with the widest variety of other images—to the particles near the center. Outliers, which are distant from the rest of the data distribution, are relegated to the boundary.

Experimental Insights

The researchers evaluated HACK on benchmarks including MNIST and CIFAR-10, as well as a synthetic "Congealed MNIST" dataset designed to test prototypicality discovery.

  • Validation via Congealing: In tests using Congealed MNIST (where digits are aligned to be perfectly typical), HACK successfully placed these aligned images at the very center of the Poincaré ball, validating the emergent prototypicality hypothesis.
  • The "Atypical" Advantage: Surprisingly, the researchers found that "atypical" images are more informative for model training. On MNIST, training a LeNet model on the 10% most atypical images resulted in a 16.54% higher accuracy than training on the 10% most typical images. These atypical samples represent the "hard examples" necessary for learning robust decision boundaries.
  • Adversarial Robustness: Conversely, typical images are better for robustness. By identifying and removing the most atypical 1% of images (extreme outliers), the authors improved a model's adversarial robustness from 84.72% to 93.42%.

Conclusion

HACK demonstrates that the shape of a feature space is as critical as the loss function used to train it. By adopting hyperbolic geometry, prototypicality is no longer an ad-hoc metric but a fundamental structural property. This discovery provides a powerful, unsupervised tool for data selection, allowing practitioners to prioritize either "hard" atypical samples for better generalization or "typical" samples for enhanced robustness.

Summary updated

Abstract

Given an image set without any labels, our goal is to train a model that maps each image to a point in a feature space such that, not only proximity indicates visual similarity, but where it is located directly encodes how prototypical the image is according to the dataset. Our key insight is to perform unsupervised feature learning in hyperbolic instead of Euclidean space, where the distance between points still reflect image similarity, and yet we gain additional capacity for representing prototypicality with the location of the point: The closer it is to the origin, the more prototypical it is. The latter property is simply emergent from optimizing the usual metric learning objective: The image similar to many training instances is best placed at the center of corresponding points in Euclidean space, but closer to the origin in hyperbolic space. We propose an unsupervised feature learning algorithm in Hyperbolic space with sphere pACKing. HACK first generates uniformly packed particles in the Poincar\'e ball of hyperbolic space and then assigns each image uniquely to each particle. Images after congealing are regarded more typical of the dataset it belongs to. With our feature mapper simply trained to spread out training instances in hyperbolic space, we observe that images move closer to the origin with congealing, validating our idea of unsupervised prototypicality discovery. We demonstrate that our data-driven prototypicality provides an easy and superior unsupervised instance selection to reduce sample complexity, increase model generalization with atypical instances and robustness with typical ones.
BibTeX
@misc{guo2023unsupervised,
  author = {Yunhui Guo and Youren Zhang and Yubei Chen and Stella X. Yu},
  title = {{Unsupervised Feature Learning with Emergent Data-Driven Prototypicality}},
  year = {2023},
  eprint = {2307.01421},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2307.01421},
}
Figure