Tiny Machine Learning Virtualization for IoT and Edge Computing using the REXA VM

Tiny Machine Learning is a new approach that is being used for data-driven prediction classification and regression on microcontrollers using local sensor data. The models are typically learned off-line and sent to the microcontroller for use as binary objects or frozen and converted static data. This approach is not universal or flexible. The REXA VM, which can virtualize embedded systems and sensor nodes and includes a general machine learning framework that supports arbitrary dynamic artificial neural network and decision tree models, is introduced in this study. The models are delivered as text files with highly compressed program code that are enclosed in code frames with embedded data (model parameters). The VM offers fundamental computations for ANN and DT models (Microservices). Using a decompiler, models can be updated (retrained) and sent to other nodes (mobile models). It can be demonstrated that virtualization using a bytecode machine and just-in-time compiler is still appropriate and effective for extremely low-resource processors.