Tricritical Phase Transition in Neural Network Learning: A New Universality Class with z = 4/3
Fuente:
Zenodo
Guardado en:
| Autor principal: | |
|---|---|
| Formato: | Recurso digital |
| Publicado: |
Zenodo
2026
|
| Materias: | |
| Acceso en línea: | |
| Etiquetas: |
Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
|
| _version_ | 1866901978749075456 |
|---|---|
| author | Clarkson, David J |
| author_facet | Clarkson, David J |
| contents | <p>We report a tricritical phase transition in neural network learning dynamics, characterised by a Binder cumulant U4* = 0.377, dynamic critical exponent z = 4/3, and a tricritical point at weight decay WD ~ 1.5 where the transition changes from continuous to first-order with bistability. These critical exponents define a new universality class (the Gradient Descent class) distinct from known equilibrium classes. The control parameter is the product WD*LR, not either alone. SGD cannot produce this transition; AdamW's adaptive learning rate is the mechanism</p> |
| format | Recurso digital |
| id | zenodo_https___doi_org_10_5281_zenodo_19141722 |
| institution | Zenodo |
| language | |
| publishDate | 2026 |
| publisher | Zenodo |
| record_format | zenodo |
| spellingShingle | Tricritical Phase Transition in Neural Network Learning: A New Universality Class with z = 4/3 Clarkson, David J universality class tricritical Binder cumulant grokking phase transition critical exponent neural network AdamW Statistical Mechanics Machine Learning <p>We report a tricritical phase transition in neural network learning dynamics, characterised by a Binder cumulant U4* = 0.377, dynamic critical exponent z = 4/3, and a tricritical point at weight decay WD ~ 1.5 where the transition changes from continuous to first-order with bistability. These critical exponents define a new universality class (the Gradient Descent class) distinct from known equilibrium classes. The control parameter is the product WD*LR, not either alone. SGD cannot produce this transition; AdamW's adaptive learning rate is the mechanism</p> |
| title | Tricritical Phase Transition in Neural Network Learning: A New Universality Class with z = 4/3 |
| topic | universality class tricritical Binder cumulant grokking phase transition critical exponent neural network AdamW Statistical Mechanics Machine Learning |
| url | https://doi.org/10.5281/zenodo.19141722 |