Neural network pruning for biomedical image segmentation

Jeong, Taehee; Bollavaram, Manasa; Delaye, Elliott; Sirasao, Ashish

doi:10.1117/12.2579256

Cited by 4 publications

(2 citation statements)

References 0 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…• https://github.com/naver/force An interesting, albeit less common, application for pruning models is within the context of segmentation. In a recent paper Jeong et al (2021) train and prune the U-Net (Ronneberger et al, 2015) architecture on two image datasets from the Cell Tracking Challenge (PhC-C2DH-U373 and DIC-C2DH-HeLa). They use the classic multi-step approach of gradually applying magnitude-pruning interleaved with fine-tuning stages.…”

Section: A3 Implementationsmentioning

confidence: 99%

“…To evaluate the flexibility of our method we used meta-gradients at the beginning of training (on a randomly initialized U-Net), prune in a single shot, and train the network once for the same number of epochs (50). We kept the training set-up the same as the baseline by Jeong et al (2021) (i.e., resizing images and segmentation maps to (256,256), setting aside 30% of training data for validation) and similarly aim to find the highest prune ratio that does not result in IOU degradation. We report intersection-over-union (IOU) metric for the two datasets in Tables 8 and 9: These results show that our method works as well (or better) compared to this compute-expensive baseline, in the sense that we can prune more parameters while keeping the IOU score the same.…”

Section: A3 Implementationsmentioning

confidence: 99%

See 1 more Smart Citation

Prospect Pruning: Finding Trainable Weights at Initialization using Meta-Gradients

Alizadeh¹,

Tailor²,

Zintgraf³

et al. 2022

Preprint

View full text Add to dashboard Cite

Pruning neural networks at initialization would enable us to find sparse models that retain the accuracy of the original network while consuming fewer computational resources for training and inference. However, current methods are insufficient to enable this optimization and lead to a large degradation in model performance. In this paper, we identify a fundamental limitation in the formulation of current methods, namely that their saliency criteria look at a single step at the start of training without taking into account the trainability of the network. While pruning iteratively and gradually has been shown to improve pruning performance, explicit consideration of the training stage that will immediately follow pruning has so far been absent from the computation of the saliency criterion. To overcome the short-sightedness of existing methods, we propose Prospect Pruning (ProsPr), which uses meta-gradients through the first few steps of optimization to determine which weights to prune. ProsPr combines an estimate of the higherorder effects of pruning on the loss and the optimization trajectory to identify the trainable sub-network. Our method achieves state-of-the-art pruning performance on a variety of vision classification tasks, with less data and in a single shot compared to existing pruning-at-initialization methods.

show abstract

Section: A3 Implementationsmentioning

confidence: 99%

Section: A3 Implementationsmentioning

confidence: 99%