KAKURENBO: Adaptively Hiding Samples in Deep Neural Network Training
This addresses efficiency issues for practitioners training large-scale models in domains like image classification and segmentation, though it is incremental as it builds on existing importance sampling methods.
The paper tackles the problem of reducing deep neural network training costs by adaptively hiding less important samples during training, achieving up to 22% reduction in training time with only a 0.4% accuracy drop compared to baseline.
This paper proposes a method for hiding the least-important samples during the training of deep neural networks to increase efficiency, i.e., to reduce the cost of training. Using information about the loss and prediction confidence during training, we adaptively find samples to exclude in a given epoch based on their contribution to the overall learning process, without significantly degrading accuracy. We explore the converge properties when accounting for the reduction in the number of SGD updates. Empirical results on various large-scale datasets and models used directly in image classification and segmentation show that while the with-replacement importance sampling algorithm performs poorly on large datasets, our method can reduce total training time by up to 22% impacting accuracy only by 0.4% compared to the baseline. Code available at https://github.com/TruongThaoNguyen/kakurenbo