Provenance Supporting Hyperparameter Analysis in Deep Neural Networks - Archive ouverte HAL Access content directly
Conference Papers Year : 2021

Provenance Supporting Hyperparameter Analysis in Deep Neural Networks

(1) , (1) , (2) , (3) , (1)
1
2
3

Abstract

The duration of the life cycle in deep neural networks (DNN) depends on the data configuration decisions that lead to success in obtaining models. Analyzing hyperparameters along the evolution of the network's execution allows for adapting the data. Provenance data derivation traces help the parameter fine-tuning by providing a global data picture with clear dependencies. Provenance can also contribute to the interpretation of models resulting from the DNN life cycle. However, there are challenges in collecting hyperparameters and in modeling the relationships between the data involved in the DNN life cycle to build a provenance database. Current approaches adopt different notions of provenance in their representation and require the execution of the DNN under a specific software framework, which limits interoperability and flexibility when choosing the DNN execution environment. This work presents a provenance data-based approach to address these challenges, proposing a collection mechanism with flexibility in the choice and representation of data to be analyzed. Experiments of the approach, using a convolutional neural network focused on image recognition, provide evidence of the flexibility, the efficiency of data collection, the analysis and the validation of network data.
Fichier principal
Vignette du fichier
IPAW_2021_Authors.pdf (940.36 Ko) Télécharger le fichier
Origin : Files produced by the author(s)

Dates and versions

lirmm-03324873 , version 1 (24-08-2021)

Identifiers

Cite

Débora Pina, Liliane Kunstmann, Daniel de Oliveira, Patrick Valduriez, Marta Mattoso. Provenance Supporting Hyperparameter Analysis in Deep Neural Networks. IPAW 2020-2021 - 8th and 9th International Provenance and Annotation Workshop, Jul 2021, London, United Kingdom. pp.20-38, ⟨10.1007/978-3-030-80960-7_2⟩. ⟨lirmm-03324873⟩
29 View
140 Download

Altmetric

Share

Gmail Facebook Twitter LinkedIn More