neural network surgery

**Neural Network Surgery** is the **practice of directly modifying a trained neural network's internal structure** — adding, removing, or reconnecting layers and neurons post-training to improve performance, efficiency, or adapt to new tasks. **What Is Neural Network Surgery?** - **Definition**: Direct manipulation of network topology or weights after initial training. - **Operations**: - **Pruning**: Remove unnecessary neurons or connections. - **Grafting**: Insert pre-trained modules from another network. - **Splicing**: Connect two networks or sub-networks together. - **Layer Removal**: Delete redundant layers (e.g., in over-deep ResNets). **Why It Matters** - **Efficiency**: Surgery can remove 90% of parameters with < 1% accuracy loss. - **Adaptation**: Quickly customize a general model for a specific deployment target. - **Debugging**: Remove or replace layers that cause specific failure modes. **Neural Network Surgery** is **precision engineering for AI** — treating trained models as modular systems that can be optimized and reconfigured post-hoc.