Can Calibration Improve Sample Prioritization?

TL;DR

This paper demonstrates that calibration techniques can enhance sample prioritization in training deep neural networks, leading to better subset selection, fewer examples per epoch, and faster overall training.

Contribution

It introduces the use of calibration methods for sample prioritization, showing they improve subset quality and training efficiency in deep learning.

Findings

Calibration improves subset quality for training

Reduces examples per epoch by at least 70%

Speeds up overall training process

Abstract

Calibration can reduce overconfident predictions of deep neural networks, but can calibration also accelerate training? In this paper, we show that it can when used to prioritize some examples for performing subset selection. We study the effect of popular calibration techniques in selecting better subsets of samples during training (also called sample prioritization) and observe that calibration can improve the quality of subsets, reduce the number of examples per epoch (by at least 70%), and can thereby speed up the overall training process. We further study the effect of using calibrated pre-trained models coupled with calibration during training to guide sample prioritization, which again seems to improve the quality of samples selected.