ReINs and RePLs: Challenging, small datasets for quick validations of designing deep neural networks for image classification
According to ImageNet's and Places365'sIntroduced 2024-07-01
- Efficiently rescaling a large dataset by adapting statistical computation to validation results of a pre-trained network.
- A unified collection of the sensitive images and those in their confused classes would form a challenging tiny set.
- An application for rescaling two large datasets: ImageNet and Places365 to obtain their rescaled subsets.
- Experimental results for image classification have validated the raised challenge of the rescaled subsets. Verifying models on these helps researchers save the computational cost and the necessary time for the early network drafts.
- It can be conducted that a network draft will obtain a good rate on large datasets if it is good on the rescaled subsets, correspondingly.