Optimizing neural network loss function with differential evolution : доклад, тезисы доклада

Описание

Тип публикации: доклад, тезисы доклада, статья из сборника материалов конференций

Конференция: IV International Conference on Advances in Materials, Earth Science and Technology (CAMSTech-IV-2024); Bukhara; Bukhara

Год издания: 2026

Идентификатор DOI: 10.1063/5.0322146

Аннотация: In this paper, a method for optimizing the parameters of the loss function for the classification problem by neural networks is proposed. A simplified Taylor polynomial was used as the loss function. Differential evolution with success rate-based adaptation was used to find the optimal parameters. The cross-entropy value was used aПоказать полностьюs the fitness value. Experiments on CIFAR-10 task can show, the purposed approach allows to achieve better results compared using Cross-Entropy. The curve of the discovered loss function exhibits a distinctly different form compared to the cross-entropy, featuring two distinct points of minimum at the probability of correct prediction equal to 0 and 1.

Ссылки на полный текст

Издание

Журнал: Proceedings of the IV International Conference on Advances in Materials, Earth Science and Technology (CAMSTech-IV-2024)

Номера страниц: 60026

Место издания: Melville

Персоны

  • Morozov Eduard (Reshetnev Siberian State University of Science and Technology)
  • Stanovov Vladimir (Siberian Federal University)
  • Gorbunov Sergey (Siberian Federal University)

Вхождение в базы данных

  • Ядро РИНЦ (eLIBRARY.RU)