Neural Network Pruning

Neural network pruning studies how to remove parts of a model that contribute little to its predictions, reducing memory use, computation, energy, or latency. Methods range from magnitude-based removal to scores using gradients, Fisher information, curvature, activations, redundancy, or task-specific behavior; they may prune individual weights or structured units such as channels, attention heads, and layers. Pruning can happen during training or after training in one shot or iteratively, often with fine-tuning, distillation, or compensation to recover lost performance. A central challenge is that the best pruning decisions depend on the model, task, calibration data, and hardware, so lower parameter counts do not always translate into faster inference or preserve reliability and reasoning behavior.