Técnicas expresivas de calibración para clasificadores multiclase
Autor (es)
Alvarez Balanya, SergioEntidad
UAM. Departamento de Tecnología Electrónica y de las ComunicacionesFecha de edición
2020-06Materias
Deep Neural Networks; Bayesian Neural Networks; Calibration; TelecomunicacionesNota
Máster Universitario en en Investigación e Innovación en Inteligencia Computacional y Sistemas InteractivosEsta obra está bajo una licencia de Creative Commons Reconocimiento-NoComercial-SinObraDerivada 4.0 Internacional.
Resumen
In the last years Deep Neural Networks (DNNs) have gained popularity as
classi cation systems and are now applied to multiple tasks where they obtain
state-of-the-art accuracy. Some of this tasks bene t from con dence
scores associated to predictions. To this aim, a probabilistic classi er outputs
con dence scores which are identi ed with the categorical distribution
over the set of possible classes. In this context, calibration is identi ed with
the validity of con dence scores as true probabilities.
Although DNNs report state-of-the-art accuracy in a plethora of tasks,
it has been pointed out recently that they present poor calibration. This
has motivated a growing body of literature on the calibration of DNNs. One
common approach, and the taken in this work, is that of training a separate
model to map uncalibrated outputs to better calibrated predictions, namely
re-calibration.
We extend the work in [1] where the authors propose to use Bayesian
Neural Networks (BNNs) as a decoupled stage from the main training of
a DNN. The Bayesian approach is appealing in the sense that it presents
proper uncertainty modeling. However, recent work [2] argues that current
approximations to the standard Bayesian approach fail to recognize
fundamental assumptions of the Bayesian paradigm. We apply the novel
framework of Generalized Variational Inference (GVI) to approximate the
posterior of the decoupled BNNs. In particular, we substitute the traditional
KL divergence used in the standard Variational Inference method for
a robust divergence. We provide a comprehensive comparison between both
divergences. Moreover, since there is no much literature on the use of robust
divergences along GVI, we also contribute with a sensitivity study in which
we analyze the in
uence of the divergence parameters
Lista de ficheros
Google Scholar:Alvarez Balanya, Sergio
Lista de colecciones del ítem
Excepto si se señala otra cosa, la licencia del ítem se describe como https://creativecommons.org/licenses/by-nc-nd/4.0/
Registros relacionados
Mostrando ítems relacionados por título, autor, creador y materia.
-
Calibración de clasificadores probabilísticos multiclase
Márquez Carrero, Sergio
2021-06 -
Generación automática de conjuntos de evaluación de camuflaje
Álvarez Balanya, Sergio
2018-07