Review & Self Study — Deployment Is a Constraint Decision

Edge, cloud, and hybrid are placement choices. None is automatically more intelligent, secure, private, cheap, or reliable.

Review questions

1. What does “edge” describe?
2. Which parts of end-to-end latency exist beyond model inference time?
3. Why can a more accurate model still be infeasible on a device?
4. What data-movement questions should replace vague privacy claims?
5. Why can a local model still depend on the network?
6. How should failure consequences influence placement?

Compare:

Edge is better because it is faster and private.

For this requirement, local inference avoids the measured network delay and keeps the specified raw sensor stream on-device, but it introduces local memory, update, and hardware-failure constraints.

The second statement exposes both evidence and cost.

Self study

• Compare the architecture of a voice assistant, robot, recommender, and photo classifier. Which likely tolerate network dependence differently?
• Find the model size of one openly documented ML model and compare it with memory on a small computing device.
• Draw a hybrid architecture with a local fallback.
• Identify one requirement that would make you reject a deployment even if its model metric were higher.