MIC: Maximizing Informational Capacity in Adaptive Representations via Isotropic Subspace Alignment
Fuente:
arXiv
Salvato in:
| Autori principali: | , , |
|---|---|
| Natura: | Preprint |
| Pubblicazione: |
2026
|
| Soggetti: | |
| Accesso online: | |
| Tags: |
Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
|
| _version_ | 1866911735722541056 |
|---|---|
| author | Hong, Dang Nguyen Nguyen, Nhi Ngoc-Yen Pham, Huy-Hieu |
| author_facet | Hong, Dang Nguyen Nguyen, Nhi Ngoc-Yen Pham, Huy-Hieu |
| contents | Although multi-scales representation learning enables elastic-dimension embeddings, nested subspaces often suffer from dimensional redundancy and spectral collapse. To address this, we introduce MIC, a framework that optimizes the geometric landscape of multi-granular embeddings through isotropic subspace alignment. MIC employs Soft Collapse Regularization (SCR) to mitigate redundancy between prefix and residual subspaces via cross-correlation penalties, alongside Spectral Isotropy Regularization (SIR) to ensure hyper-spherical uniformity in low-dimensional prefixes. By unifying these strategies through a self-distillation objective, MIC generates semantically dense representations that maintain high discriminative power. Our experiments demonstrate that MIC significantly outperforms standard baselines, particularly in high-compression scenarios where maintaining informational capacity is most critical. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2605_29987 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | MIC: Maximizing Informational Capacity in Adaptive Representations via Isotropic Subspace Alignment Hong, Dang Nguyen Nguyen, Nhi Ngoc-Yen Pham, Huy-Hieu Machine Learning Computation and Language Although multi-scales representation learning enables elastic-dimension embeddings, nested subspaces often suffer from dimensional redundancy and spectral collapse. To address this, we introduce MIC, a framework that optimizes the geometric landscape of multi-granular embeddings through isotropic subspace alignment. MIC employs Soft Collapse Regularization (SCR) to mitigate redundancy between prefix and residual subspaces via cross-correlation penalties, alongside Spectral Isotropy Regularization (SIR) to ensure hyper-spherical uniformity in low-dimensional prefixes. By unifying these strategies through a self-distillation objective, MIC generates semantically dense representations that maintain high discriminative power. Our experiments demonstrate that MIC significantly outperforms standard baselines, particularly in high-compression scenarios where maintaining informational capacity is most critical. |
| title | MIC: Maximizing Informational Capacity in Adaptive Representations via Isotropic Subspace Alignment |
| topic | Machine Learning Computation and Language |
| url | https://arxiv.org/abs/2605.29987 |