Are All Training Examples Created Equal? An Empirical Study
arXiv:1811.12569
Abstract
Modern computer vision algorithms often rely on very large training datasets. However, it is conceivable that a carefully selected subsample of the dataset is sufficient for training. In this paper, we propose a gradient-based importance measure that we use to empirically analyze relative importance of training images in four datasets of varying complexity. We find that in some cases, a small subsample is indeed sufficient for training. For other datasets, however, the relative differences in importance are negligible. These results have important implications for active learning on deep networks. Additionally, our analysis method can be used as a general tool to better understand diversity of training examples in datasets.
arXiv comment12 pages, 12 figures
References in corpus (4)
- Very Deep Convolutional Networks for Large-Scale Image Recognition
- Man is to Computer Programmer as Woman is to Homemaker? Debiasing Word Embeddings
- No Classification without Representation: Assessing Geodiversity Issues in Open Data Sets for the Developing World
- Adversarial Active Learning for Deep Networks: a Margin Based Approach
Cited by in corpus (13)
- Are all negatives created equal in contrastive instance discrimination?
- Learning from others' mistakes: Avoiding dataset biases without modeling them
- Semantic Redundancies in Image-Classification Datasets: The 10% You Don't Need
- Efficient Visual Recognition with Deep Neural Networks: A Survey on Recent Advances and New Directions
- A Study of Gradient Variance in Deep Learning
- EcoNAS: Finding Proxies for Economical Neural Architecture Search
- Representation Matters: Assessing the Importance of Subgroup Allocations in Training Data
- Sampling Bias in Deep Active Classification: An Empirical Study
- Algorithm Fairness in AI for Medicine and Healthcare
- One Backward from Ten Forward, Subsampling for Large-Scale Deep Learning
- Using Wavelets to Analyze Similarities in Image-Classification Datasets
- Adapting Convolutional Neural Networks for Geographical Domain Shift
- Training Efficiency and Robustness in Deep Learning