Feedback Control for Online Training of Neural Networks

Zilong Zhao 1 Sophie Cerf 1 Bogdan Robu 1 Nicolas Marchand 1
1 GIPSA-SYSCO - SYSCO
GIPSA-DA - Département Automatique
Abstract : Convolutional neural networks (CNNs) are commonly used for image classification tasks, raising the challenge of their application on data flows. During their training, adaptation is often performed by tuning the learning rate. Usual learning rate strategies are time-based i.e. monotonously decreasing. In this paper, we advocate switching to a performance- based adaptation, in order to improve the learning efficiency. We present E/PI-Control, a conditional learning rate strategy that combines a feedback PI controller based on the CNN loss function, with an exponential control signal to smartly boost the learning and adapt the PI parameters. Stability proof is provided as well as an experimental evaluation using two state of the art image datasets (CIFAR-10 and Fashion-MNIST). Results show better performances than the related works (faster network accuracy growth reaching higher levels) and robustness of the E/PI-Control regarding its parametrization.
Complete list of metadatas

https://hal.archives-ouvertes.fr/hal-02115916
Contributor : Zilong Zhao <>
Submitted on : Tuesday, April 30, 2019 - 3:34:29 PM
Last modification on : Thursday, May 16, 2019 - 11:55:06 AM

Identifiers

  • HAL Id : hal-02115916, version 1

Collections

Citation

Zilong Zhao, Sophie Cerf, Bogdan Robu, Nicolas Marchand. Feedback Control for Online Training of Neural Networks. IEEE Conference on Control Technology and Applications, Aug 2019, Hong Kong, China. ⟨hal-02115916⟩

Share

Metrics

Record views

36