Running edge devices with TinyML
After the model is trained using the federated learning approach that we have discussed so far, we want to deploy the trained model and conduct efficient model inference/serving. This leads to the concept of TinyML.
The deploy hardware of edge devices usually has a lot of constraints. Let's look at these constraints and how we can tackle them:
- Limited battery power: This means that our deployment should be very efficient and cannot consume a lot of battery power.
- Unstable connection to the server: This means that we need to guarantee that the model is still usable if the device cannot connect to the server.
- High latency for communication: This means that if some emergency happens, the model deployed on the device can handle it without coordinating with the central server.
- Data locality: This means that we need to keep each device's local data private and never allow the local data to communicate with other devices. ...
