SECAI Core · phase 2 of 15
Pruning
Removing less important elements of the model that have negligible impact on performance
The Explain card
- Plain English
- Pruning removes parts of a model, such as weights or neurons, that contribute almost nothing to its output, making it smaller and faster with little loss in performance.
- Example
- An engineer prunes a vision model by half so it fits on an edge camera, then re-tests accuracy to confirm detection still works.
- Why it matters
- Pruning changes behaviour in subtle ways. It can strip safety-relevant capacity, and some attacks design a backdoor that only activates after a model is pruned or compressed.
- Hook
- Trim the dead branches, but check what was living in them.
Where it sits in the deck
Phase 2: How Models Are Built: Architectures and Learning Mechanics
With the paradigms named, zoom in on the architectural building blocks that turn data into trained artifacts.
