Vers un désenchevêtrement de l'ambiguïté de la tâche et de l'incertitude du modèle pour la classification avec option de rejet à l'aide de réseaux neuronaux

Titouan Lorieul 1 Alexis Joly 1
1 ZENITH - Scientific Data Management
LIRMM - Laboratoire d'Informatique de Robotique et de Microélectronique de Montpellier, CRISAM - Inria Sophia Antipolis - Méditerranée
Abstract : Classification with reject option is a way to address the problem of estimating the uncertainty of a classifier. Recent approaches to this problem use criteria based on either a confidence or a dispersion measure. However, they do not explicitly combine the two main sources of uncertainty : the ambiguity of the task, inherent to it, and the uncertainty of the model, resulting from data sampling and stochasticity of learning process. In this article, we explore how these two quantities can be merged to build more effective rejection criteria. In particular, we propose methods for combining disagreement measures and ambiguity estimates using an ensemble of models. Experiments on synthetic data sets constructed to model different types of uncertainties indicate that these new criteria have similar performance to the baselines. Nevertheless, more in-depth analyses show empirical evidence that highlights the existence of additional information in the distribution of the overall results. In practice, the ideal rejector may be a more complex function than the previous criteria, and may even be counter-intuitive at times.
Complete list of metadatas

Cited literature [25 references]  Display  Hide  Download

https://hal.archives-ouvertes.fr/hal-02421210
Contributor : Titouan Lorieul <>
Submitted on : Friday, December 20, 2019 - 12:25:39 PM
Last modification on : Tuesday, January 21, 2020 - 3:35:24 PM

File

article_CAP_2019.pdf
Files produced by the author(s)

Identifiers

  • HAL Id : hal-02421210, version 1

Collections

Citation

Titouan Lorieul, Alexis Joly. Vers un désenchevêtrement de l'ambiguïté de la tâche et de l'incertitude du modèle pour la classification avec option de rejet à l'aide de réseaux neuronaux. CAp 2019 - Conférence sur l'Apprentissage automatique, Jul 2019, Toulouse, France. ⟨hal-02421210⟩

Share

Metrics

Record views

17

Files downloads

15