Title: STARFISH: Efficient Accuracy Recovery in Pruned Networks via Internal State Healing

Abstract:

Network pruning is a technique employed to decrease the weight count within large neural architectures. While this approach significantly accelerates inference, it often leads to a notable drop in model precision, necessitating a subsequent healing phase to restore performance. This study introduces STARFISH, a novel healing framework capable of efficiently restoring (the majority of) the accuracy in any pruned network. The core concept behind STARFISH involves optimizing the pruned structure to align its internal state representations with those of the original, unpruned network, utilizing a minimal calibration set composed of unlabeled samples. In scenarios involving the removal of 50% of weights, STARFISH enhances accuracy recovery by as much as 22% compared to current state-of-the-art methods, particularly on Vision Transformer (ViT) based models. This benefit becomes even more significant under aggressive pruning conditions. For instance, when 75% of weights are eliminated from a DeiT-B network for ImageNet classification, STARFISH requires only 0.4% of the standard training dataset size for calibration, successfully recovering 82% of the original dense model’s accuracy. In contrast, alternative recovery techniques achieve merely 40% of the dense model’s performance under similar conditions.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC