Pruning Ratio

Pruning ratio, the proportion of parameters or data points removed from a model, is a critical factor in optimizing machine learning models for efficiency without significant performance loss. Current research focuses on developing algorithms that automatically determine optimal pruning ratios for various architectures, including deep neural networks and 3D Gaussian splatting, often leveraging techniques like combining magnitude and relevance scores or employing dynamic programming for layer-adaptive pruning. These advancements aim to improve model compression, reduce computational costs, and enhance resource-constrained applications while maintaining or even improving accuracy, particularly addressing challenges like class imbalance and the impact of pruning on different model layers.