EURO-Online login
- New to EURO? Create an account
- I forgot my username and/or my password.
- Help with cookies
(important for IE8 users)
2211. A Novel Criterion for Batch Size Adaptation in Stochastic Gradient Methods
Invited abstract in session TD-3: Data science meets strongly NP-Hard CO , stream Data Science Meets Optimization.
Tuesday, 14:30-16:00Room: 1005 (building: 202)
Authors (first author is the speaker)
| 1. | Marco Boresta
|
| Istituto di analisi dei sistemi ed informatica "Antonio Ruberti" (IASI), Consiglio Nazionale delle Ricerche (CNR) | |
| 2. | Alberto De Santis
|
| Sapienza University of Rome | |
| 3. | Stefano Lucidi
|
| Department of Computer, Control, and Management Science, University of Rome "La Sapienza" |
Abstract
In this work, we introduce a novel interpretation of the variance associated with batch gradient estimates in machine learning model training. By reevaluating this variance, we introduce a novel strategy for the dynamic adjustment of batch sizes throughout the training process. This approach diverges from traditional approaches by prioritizing a more detailed representation of variance behavior, which can lead to more efficient and effective training regimes. We test our methodology against a series of problems, benchmarking it against current state-of-the-art methods to demonstrate its advantages and applicability.
Keywords
- Analytics and Data Science
- Machine Learning
Status: accepted
Back to the list of papers