Efficient on-device learning for resource-constrained devices

[christophe0606]

Experiment with an optimized on-device machine learning framework that could be used to:

• Fine-tune the last layers of a network, with the preceding layers computed using Ethos or CMSIS-NN.
• Train adaptive filters.
• Support other lightweight on-device learning use cases.

The framework leverages the work done on the CMSIS-DSP C++ extension and its loop-fusion technology to implement efficient backward passes.

The framework focuses on:

• Fine-tuning.
• Linear algebra operations required for training.
• Efficient execution on resource-constrained devices.

Fine-tuning often involves fully connected layers, so the framework is intentionally focused on the operations needed to train these layers efficiently.

This framework is not intended to be a full-featured machine learning framework. In particular, it will not provide layers such as convolutions or pooling. These operations are expected to be handled by existing inference frameworks such as CMSIS-NN or NPU like Ethos.