On the pitfalls of learning with limited data: A facial expression recognition case study
作者：
Highlights：
• High variance data trained with prior information perform better on stable test data.
• Data with fewer changes translate to more stable performance on novel scenarios.
• Only merging heterogeneous data is not a straightforward improvement.
• Classical data augmentation cannot fill the holes from joining separated datasets.
• Inductive biases and synthetic data help to bridge the gap among diverse datasets.
摘要
•High variance data trained with prior information perform better on stable test data.•Data with fewer changes translate to more stable performance on novel scenarios.•Only merging heterogeneous data is not a straightforward improvement.•Classical data augmentation cannot fill the holes from joining separated datasets.•Inductive biases and synthetic data help to bridge the gap among diverse datasets.
论文关键词：Learning with limited data,Limited data,Video classification
论文评审过程：Received 31 August 2020, Revised 23 February 2021, Accepted 1 April 2021, Available online 23 April 2021, Version of Record 30 June 2021.
论文官网地址：https://doi.org/10.1016/j.eswa.2021.114991
