Comparison of Optimization Algorithms Used in Deep Learning by Using Caltech 101 Data Set


Creative Commons License

Seyyarer E., Ayata F., Uçkan T., Hark C., Karcı A.

International Conference on Data Science, Machine Learning and Statistics - 2019 (DMS-2019), Van, Turkey, 26 June 2017 - 29 June 2019, pp.18

  • Publication Type: Conference Paper / Summary Text
  • City: Van
  • Country: Turkey
  • Page Numbers: pp.18
  • Van Yüzüncü Yıl University Affiliated: Yes

Abstract

In our previous study, an artificial neural network was applied to the caltech 101 data set which is an international data set, the results were analyzed and converted into a publication. In this study, image preprocessing, segmentation and feature selection were performed. 7 invariant moments applied to geometric and colorless images were used for feature selection. The success rate in the classification was observed to be about 25%. In this study, deep learning and optimization techniques were applied on the same data set. Relu was used as activation function and cross entropy was preferred as loss function. Images are resized to 64x64. Each time the program is run, a random 6-category image is taken and 100 iterations are executed. Stochastic gradient descent (sgd), momentum, adam, adagrad, rmsprop and adadelta optimization algorithms were used for different results and these results were analyzed. The success rates in the classification were as follows: sgd: 64.5%, momentum: 85.56%, adam: 92.31%, adagrad: 71.25%, rmsprop: 40.26% and adadelta: 86.88%.