TL;DR
This paper introduces a framework to successfully apply data augmentation to Deep Subspace Clustering (DSC). By combining a novel subspace consistency loss with temporal ensembling and an unsupervised greedy search for augmentation policies, the authors achieve state-of-the-art clustering performance across multiple standard datasets.
The Challenge of Unsupervised Augmentation
In supervised learning, data augmentation is a "free lunch" for improving generalization; we know that a rotated image of a cat is still a cat because the label tells us so. However, in Deep Subspace Clustering (DSC)—where the goal is to group high-dimensional data points (like faces or objects) into their respective low-dimensional subspaces without any labels—applying augmentation is tricky.
Without ground-truth labels, how do you tell a neural network that a transformed version of an image should still belong to the same cluster? Standard DSC networks are prone to overfitting and lack a mechanism to ensure that "slightly changed" inputs yield "consistent" clustering results. This paper bridges that gap by enforcing representational consistency.
Methods: Consistency and Policy Search
The authors propose a two-pronged approach to integrate data augmentation into DSC effectively:
1. Subspace Membership Consistency Loss
The core idea is that transformed versions of the same data point should maintain the same subspace membership in the latent space.
To implement this, the authors introduce a consistency loss based on temporal ensembling. Because augmentations are stochastic, the network's internal representations can fluctuate wildly. To stabilize this, they maintain an Exponential Moving Average (EMA) of the coefficient matrices (which represent subspace relationships) from previous training iterations. This "ensemble" serves as a stable target, penalizing the network if the current augmented input produces a subspace prediction that deviates too far from the historical average.
2. Unsupervised Augmentation Policy Search
Not all augmentations are helpful (e.g., flipping a digit '6' might turn it into a '9'). Since they have no labels to validate a policy, the authors use a greedy search algorithm. They test various image operations (rotation, shear, contrast, etc.) and evaluate their effectiveness using the mean Silhouette coefficient—an internal clustering metric that measures how well-separated the resulting clusters are. This allows the model to "self-select" the best augmentations for a specific dataset without human intervention.
Datasets
The method was evaluated on four standard subspace clustering benchmarks, which feature varying degrees of complexity, lighting changes, and object poses:
- Extended Yale-B: 2,432 facial images of 38 individuals under varying lighting.
- ORL: 400 facial images of 40 individuals.
- COIL-20 & COIL-100: 20 and 100 objects, respectively, photographed at different poses on a turntable.
Key Insights and Results
The proposed method, MLRDSC-DA, achieved state-of-the-art (SOTA) results, significantly outperforming previous methods:
- Massive Error Reduction: On the Extended Yale B dataset, the method reduced clustering error to 0.82% (down from the previous SOTA of 1.36%). On COIL-20, it achieved an error of 1.79% (down from 2.08%).
- Consistency is Mandatory: An ablation study revealed a critical insight: data augmentation alone can actually degrade performance in clustering if the network isn't explicitly told to keep the outputs consistent. For example, on Yale B, using augmentations without the consistency loss actually increased the error from 1.36% to 2.84%. The consistency loss is the "glue" that makes augmentation work.
- Labels are Optional for Policy Search: One of the most surprising findings is that the Silhouette coefficient—a purely geometric measure of cluster quality—is a robust enough proxy to find augmentation policies that generalize well to the actual (hidden) ground-truth labels.
Limitations and Future Work
While highly effective, the approach has a few limitations. First, the greedy search, though more efficient than brute force, still requires training "child" networks to evaluate policies, which can be computationally expensive for very large datasets. Second, the search is limited to a discrete set of common image processing operations; future work could explore learning "neural" augmentations specifically tailored for subspace structures. Finally, like most DSC methods, this assumes data lies in a union of linear subspaces in a latent space, which remains a geometric constraint for extremely complex data distributions.
Summary updated