Implicit Hypothesis Testing and Divergence Preservation in Neural Network Representations
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_ | 1866909045039824896 |
|---|---|
| author | Aksoy, Kadircan Bhattacharjee, Protim Jung, Peter |
| author_facet | Aksoy, Kadircan Bhattacharjee, Protim Jung, Peter |
| contents | We study the training dynamics of neural classifiers through the lens of binary hypothesis testing. We re-formalize classification as a collection of binary tests between class-conditional distributions induced by learned representations and show empirically that, along training trajectories, well-generalizing networks progressively approach Neyman-Pearson optimal decision rules, as measured by monotonic growth in the KL divergence retained by learned representations. We provide sufficient conditions for exact optimality, discuss its implications for training regularization, and define an informational plane, (so-called Evidence-Error plane) where convergence can be assessed methodically across network architecture. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2601_20477 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | Implicit Hypothesis Testing and Divergence Preservation in Neural Network Representations Aksoy, Kadircan Bhattacharjee, Protim Jung, Peter Machine Learning Information Theory We study the training dynamics of neural classifiers through the lens of binary hypothesis testing. We re-formalize classification as a collection of binary tests between class-conditional distributions induced by learned representations and show empirically that, along training trajectories, well-generalizing networks progressively approach Neyman-Pearson optimal decision rules, as measured by monotonic growth in the KL divergence retained by learned representations. We provide sufficient conditions for exact optimality, discuss its implications for training regularization, and define an informational plane, (so-called Evidence-Error plane) where convergence can be assessed methodically across network architecture. |
| title | Implicit Hypothesis Testing and Divergence Preservation in Neural Network Representations |
| topic | Machine Learning Information Theory |
| url | https://arxiv.org/abs/2601.20477 |