TopoPrune: Robust Data Pruning via Unified Latent Space Topology

TopoPrune introduces a topology-based framework for data pruning that addresses instability issues in geometric methods by leveraging intrinsic data structure rather than extrinsic geometry. The approach combines manifold approximation with persistent homology to achieve high accuracy at extreme pruning rates (90%) while maintaining robustness across architectures and noise conditions.

TopoPrune represents a meaningful advance in machine learning efficiency by tackling a fundamental limitation of existing geometric pruning methods. Current approaches rely on extrinsic geometric properties that prove brittle when latent spaces shift, limiting their practical deployment across different neural network architectures. This instability has constrained the real-world applicability of data pruning despite its theoretical promise for reducing computational overhead.

The research builds on growing recognition that topological properties offer more stable representations than geometric ones. By operating at two complementary scales—establishing a global low-dimensional embedding through topology-aware manifold approximation, then performing local optimization via persistent homology—TopoPrune captures both macro and micro structural patterns. This dual-scale approach mirrors successful strategies in other machine learning domains where multi-level analysis improves robustness.

The practical implications are substantial for resource-constrained environments. Achieving 90% pruning rates with maintained accuracy directly translates to reduced training costs, lower inference latency, and decreased model deployment footprints. The demonstrated transferability across architectures addresses a critical pain point: models trained on pruned datasets using one architecture can generalize effectively to others, enabling broader adoption of data-efficient learning practices.

The noise resilience property carries particular significance for production environments where feature embeddings naturally drift. TopoPrune's intrinsic stability suggests practitioners can implement this method with greater confidence than existing alternatives. Going forward, the framework's principled topological foundation could inspire similar approaches in other pruning domains—weight pruning, layer pruning, or hybrid strategies—potentially establishing topology as a foundational paradigm for model compression.

• →TopoPrune achieves stable data pruning at 90% rates by using topological properties instead of fragile geometric features.
• →The dual-scale approach combines global manifold approximation with local persistent homology analysis for comprehensive data structure capture.
• →Framework demonstrates exceptional robustness to latent space noise and superior cross-architecture transferability compared to existing methods.
• →Topology-based approach offers principled foundation that may inspire broader adoption across weight pruning and other compression techniques.
• →Practical implications include reduced training costs, faster inference, and smaller deployment footprints for resource-constrained environments.