The model training process in Neuton contains the following steps:

A Virtual Machine (VM) is automatically provisioned to perform the following tasks: digital signal processing, model training, validation, and preparation of the C Library for inference. This process is fully automated and does not require any user input or action. Creating a virtual machine may take several minutes and is not included in the usage cost.

Digital Signal Processing (DSP)

This option makes data suitable for creating TinyML models. DSP consists of windowing, feature extraction, and feature selection. The user can flexibly choose the needed settings.

During this step, our proprietary neural network algorithm, Neuton, will automatically create a neural network architecture to achieve the best possible score on the validation data (measured by the metric selected by the user).

Once model training is complete, a downloadable archive is generated. It contains everything you need to embed the model into an MCU or programmable sensor. It is also possible validate the results of the inference of the model even without implementation in the Edge device using Neuton Inference Runner Executable.

Upon completion of training, the virtual machine will be automatically deprovisioned.