Embedded AI
Learning paradigms: how the model learns
The right way to train depends on the data you have. Labelled examples, only normal operation, or nothing yet: each case leads to a different approach.
· by BCF Embedded engineering team
Approaches we use
Supervised learning
Learns from labelled examples. Best when the classes are known and data can be labelled.
Unsupervised learning
No labels needed. Clustering and autoencoders learn normal behaviour and detect anomalies.
Self- and semi-supervised
Lots of unlabelled data plus a small labelled set, to cut the cost of labelling.
Transfer learning
A pre-trained model fine-tuned on a small dataset from your device.
Reinforcement learning
Learns a control policy from rewards, usually in simulation, then runs as a fixed policy.
On-device learning
The model adapts on the device: calibration to one machine or user, or slow drift.
Trained in the lab or adapting on the device
Trained offline, deployed frozen
- Predictable behaviour, easy to validate
- Smallest footprint on the device
- Improved versions shipped over secure OTA
Adapting on the device
- Calibration to each unit or user
- Handles drift of sensors and machines
- Raw data stays on the device
Federated learning
When data cannot leave the devices, a fleet can train a shared model by sending model updates instead of raw data. It adds work in the firmware, communication and validation, so we use it only where privacy or bandwidth requires it.
FAQ
Discuss your project
Describe your device and goal. An engineer replies within one business day.