What are the trade-offs between model accuracy and inference latency, and how do you optimize for the latter on edge devices?
