New Pruning Method Preserves Model Calibration and Reliability

Model pruning is a common technique to reduce the size of neural networks, but ensuring the reliability of the pruned model's predictions, especially in terms of calibration, remains a challenge. This research proposes Calibration-Preserving Pruning (CPP), which aims to achieve smaller, yet still valid, prediction sets after compression. The method enhances a standard pruning score by incorporating nonconformity-gradient saliency, which helps identify parameters crucial for maintaining prediction reliability. CPP employs a rigorous data splitting strategy, using separate sets for pruning, validation, conformal calibration, and testing. This ensures that the finite-sample marginal coverage guarantee from split conformal prediction remains valid. Experimental results on Qwen2.5-1.5B models, pruned to 50% sparsity, show that CPP-SparseGPT significantly reduces mean prediction set size and often improves accuracy on large-label tasks like DBpedia-14. While generic supervised gradients contribute to gains, CPP specifically targets and improves reliability-sensitive classification.