In the present work we present Training Noise Token (TNT) Pruning for vision transformers. Our method relaxes the discrete token dropping condition to continuous additive noise, providing smooth optimization in training, while retaining discrete dropping computational gains in deployment settings. We provide theoretical connections to Rate-Distortion literature, and empirical evaluations on the ImageNet dataset using ViT and DeiT architectures demonstrating TNT's advantages over previous pruning methods.
View on arXiv@article{rao2025_2411.18092,
title={ Training Noise Token Pruning },
author={ Mingxing Rao and Bohan Jiang and Daniel Moyer },
journal={arXiv preprint arXiv:2411.18092},
year={ 2025 }
}