Pruning exploits the fact that trained neural networks are usually over-parameterized: many weights are close to zero and can be removed. Structured pruning deletes entire filters or channels, which directly reduces compute on any hardware; unstructured pruning zeroes individual weights and only helps when the runtime supports sparse operations.
For IoT devices the goal is to fit the model into flash and RAM and to shorten inference time so the device can sleep sooner. The trade-off is engineering effort: pruned models usually need fine-tuning to recover accuracy, and aggressive pruning can quietly degrade performance on rare classes. Pruning is typically paired with quantization when preparing a model for a microcontroller.
Part of the IoT Glossary. Spotted an error or a missing term? Tell us.