Mostrar el registro sencillo de la publicación
Quantized SG-MCMC for Bayesian deep posterior compression
| dc.contributor.author | Hernández, Sergio | |
| dc.contributor.author | López-Cortes, Xaviera | |
| dc.date.accessioned | 2025-05-29T18:59:14Z | |
| dc.date.available | 2025-05-29T18:59:14Z | |
| dc.date.issued | 2025 | |
| dc.identifier.uri | http://repositorio.ucm.cl/handle/ucm/6064 | |
| dc.description.abstract | In this paper, we propose a novel quantization technique for Bayesian deep learning aimed at enhancing efficiency without compromising performance. Our approach leverages post-training quantization to significantly reduce the memory footprint of stochastic gradient samplers, particularly Stochastic Gradient Markov Chain Monte Carlo (SG-MCMC) methods. This technique achieves a level of compression comparable to optimal thinning, which traditionally necessitates not only the original samples in single precision floating-point representation but also the gradients, resulting in substantial computational overhead. In contrast, our quantization method requires only the original samples and can accurately recover posterior modes through a simple affine transformation. This process incurs minimal additional memory or computational costs, making it a highly efficient alternative for Bayesian deep learning applications. | es_CL |
| dc.language.iso | en | es_CL |
| dc.rights | Atribución-NoComercial-SinDerivadas 3.0 Chile | * |
| dc.rights.uri | http://creativecommons.org/licenses/by-nc-nd/3.0/cl/ | * |
| dc.source | Communications in Computer and Information Science, 2270, 158-169 | es_CL |
| dc.title | Quantized SG-MCMC for Bayesian deep posterior compression | es_CL |
| dc.type | Article | es_CL |
| dc.ucm.facultad | Facultad de Ciencias de la Ingeniería | es_CL |
| dc.ucm.indexacion | Scopus | es_CL |
| dc.ucm.uri | springerlink.ucm.elogim.com/chapter/10.1007/978-3-031-80084-9_11 | es_CL |
| dc.ucm.doi | doi.org/10.1007/978-3-031-80084-9_11 | es_CL |
Ficheros en la publicación
| Ficheros | Tamaño | Formato | Ver |
|---|---|---|---|
|
No hay ficheros asociados a esta publicación. |
|||
Esta publicación aparece en la(s) siguiente(s) colección(ones)
Excepto si se señala otra cosa, la licencia de la publicación se describe como Atribución-NoComercial-SinDerivadas 3.0 Chile