State Representation Learning Using an Unbalanced Atlas

Li Meng, Morten Goodwin, Anis Yazidi, Paal Engelstad

Abstract

Self-supervised learning (SSL) frameworks based on the manifold hypothesis often struggle with scaling due to "head collapse" caused by uniform distribution constraints. We propose the Unbalanced Atlas (UA), a paradigm that uses Maximum Mean Discrepancy (MMD) to encourage non-uniform chart usage, thereby increasing model confidence and performance at high dimensions. Our method, DIM-UA, demonstrates significant improvements over the ST-DIM baseline on Atari benchmarks, particularly as model capacity scales.

Introduction

The manifold hypothesis suggests that high-dimensional data, such as images, lies on a lower-dimensional manifold. Learning this structure via an "atlas"—a set of local maps or "charts"—is a promising direction for SSL. However, existing methods often enforce a uniform distribution of data across these charts, leading to high prediction uncertainty and redundant representations. This "head collapse" has historically limited manifold-based SSL to very small hidden dimensions, making them impractical for modern deep learning architectures.

The Unbalanced Atlas Approach

The Unbalanced Atlas (UA) deliberately avoids uniform distributions. By employing an MMD loss, the model is pushed away from uniformity, allowing it to become more certain about which chart best represents specific data points. This reduction in entropy is a key driver of the model's performance.

Entropy comparison showing lower uncertainty in the Unbalanced Atlas compared to a uniform prior.

To maintain diversity and prevent charts from converging on the same solution, the authors use "dilated prediction targets" based on Minkowski sums. This ensures that information is preserved even in overlapping regions of the manifold, preventing the "boring" solution where all heads output identical information.

Experimental Results

The UA paradigm was integrated into the Spatiotemporal DeepInfomax (ST-DIM) framework, resulting in DIM-UA. Evaluation on the Atari Annotated RAM Interface (AtariARI) and CIFAR-10 benchmarks revealed that UA effectively breaks the "dimensionality ceiling" that previously limited manifold learning.

Performance comparison of DIM-UA vs. ST-DIM across varying hidden unit sizes, demonstrating superior scaling.

As shown in the scaling analysis, DIM-UA continues to improve as the number of hidden units increases, whereas the baseline ST-DIM performance plateaus or even collapses at high capacities. On AtariARI, DIM-UA achieved a mean F1 score of ~75%, a significant jump from the ~70% baseline at 16,384 units. Notably, in games like Freeway, DIM-UA remained stable (F1 0.86) while the standard ST-DIM model collapsed (F1 0.3).

Conclusion and Future Work

The Unbalanced Atlas provides a scalable path for manifold-based representation learning. By embracing imbalance, the model achieves higher confidence and better utilizes increased capacity. While effective, the approach relies on certain mathematical assumptions regarding chart convexity and exhibits sensitivity to hyperparameters like the MMD loss weight. Future research will focus on finding optimal head counts based on intrinsic data dimensionality and developing more efficient diversity-maintaining loss functions.

Summary updated

BibTeX
@inproceedings{meng2024state,
  author = {Li Meng and Morten Goodwin and Anis Yazidi and Paal Engelstad},
  title = {{State Representation Learning Using an Unbalanced Atlas}},
  booktitle = {International Conference on Learning Representations},
  year = {2024},
  url = {https://openreview.net/forum?id=cWdAYDLmPa},
}
Figure