Automatic differentiation, or autodiff, helps an algorithm learn from its errors. You describe a calculation and a measure of how far its result is from the desired result. Autodiff then calculates how each adjustable parameter affects that error. An optimizer uses this information to adjust the parameters and try to reduce the error. You do not have to work out the derivative formulas by hand.

CMSIS-DSP autodiff is experimental. Its focus is on-device fine-tuning using existing CMSIS-DSP kernels. It is not a new machine-learning framework.

When it helps

Autodiff can help when an algorithm needs to adapt to new measurements or to a particular device. For example, you might learn filter coefficients from a reference signal, or fine-tune the last layer of a classifier using new examples. The aim is to adjust selected parameters in a small computation, without retraining an entire model.

Why it uses the C++ API

The DSP++ C++ API provides a higher-level way to combine existing CMSIS-DSP kernels into new algorithms while keeping good performance. Its loop fusion feature can combine several operations into a single loop, avoiding temporary arrays and repeated passes over the data. Building autodiff on this API makes it easier to write efficient fine-tuning algorithms from the existing kernels. See the DSP++ introduction for an example of how this works.

See the Iris classification example for a complete training and inference application.