ImageDataset2Vec: An image dataset embedding for algorithm selection
作者：
Highlights：
• Method extracts a vectorial embedding representation of image classification tasks.
• Method evaluated in a metalearning case study with six CNN algorithms and 45 datasets.
• Feature vectors of datasets with similar difficulty are close in the embedding space.
• The results overcame the baseline methods in all assessed performance measures.
• Meta-learning with Dataset2Vec reached high precision, overcoming image descriptors.
摘要
•Method extracts a vectorial embedding representation of image classification tasks.•Method evaluated in a metalearning case study with six CNN algorithms and 45 datasets.•Feature vectors of datasets with similar difficulty are close in the embedding space.•The results overcame the baseline methods in all assessed performance measures.•Meta-learning with Dataset2Vec reached high precision, overcoming image descriptors.
论文关键词：Dataset embedding,Algorithm selection,Meta-learning,Image dataset description
论文评审过程：Received 21 August 2020, Revised 21 December 2020, Accepted 14 April 2021, Available online 23 April 2021, Version of Record 6 May 2021.
论文官网地址：https://doi.org/10.1016/j.eswa.2021.115053
