1. Home

OPTIMAL DATASET SELECTION FOR TRANSFERLEARNING

Sandis Deksnis, Rolands Piterāns, Sergejs Kodors

Abstract


The proposed article describes transfer learning and Earth Mover’s Distance (EMD) methodology application in machine learning. The goal was to find out the shortest distance among three datasets in order to identifyt dataset, which is more suited for neural network pretraining. The experiment was completed using Python programming language and Jupyter Notebook. Neural network pretrained on ImageNet dataset was applied as feature extractor. The extracted feature vectors of datasets were applied to calculate the minimal distance using EMD algorithm.

Keywords


datasets; Earth Mover’s Distance (EMD); ImageNet; neural networks; transfer learning;

Full Text:

PDF

References


Chi Zhang, Yujun Cai, Guosheng Lin, Chunhua Shen “DeepEMD: Differentiable Earth Mover’s Distance for Few-Shot Learning” Sk. internetā. (2020.) - https://arxiv.org/pdf/2003.06777.pdf

Haris Pozidis, Kubilay Atasu “Linear-Complexity Earth Mover’s Distance Approximations for Efficient Similarity Search” Sk. internet (07.16.2019.) - https://www.ibm.com/blogs/research/2019/07/earth-movers-distance/

Rohan Saha, Debaruna Saha “Transfer Learning – A Comparative Analysis” Sk. internetā. (2018.) - https://www.researchgate.net/publication/329786975_Transfer_Learning_-_A_Comparative_Analysis

Yin Cui, Yang Song, Chen Sun, Andrew Howard, Serge Belongie “Large Scale Fine-Grained Categorization and Domain-Specific Transfer Learning” (arXiv:1806.06193v1) Sk. internetā. (2018.) - https://openaccess.thecvf.com/content_cvpr_2018/papers/Cui_Large_Scale_Fine-Grained_CVPR_2018_paper.pdf

Dataset “Plants_Dataset[99 classes]” by Muhammad jawad - https://www.kaggle.com/muhammadjawad1998/plants-dataset99-classes

Dataset “Animals-10” by Corrado Alessio - https://www.kaggle.com/alessiocorrado99/animals10

Dataset “Flowers Recognition” by Alexander Mamaev - https://www.kaggle.com/alxmamaev/flowers-recognition.




DOI: https://doi.org/10.17770/het2021.25.6777

Refbacks

  • There are currently no refbacks.
Rēzeknes Tehnoloģiju akadēmija, 2016 © Atbrīvošanas aleja 115, Rēzekne, Latvija. Tālrunis: 28325368;
Jautājumus un ierosinājumus sūtiet uz helpdesk@rta.lv