AI on Microcontrollers

Executing neural networks on embedded processors like Cortex-M4 presents several challenges: limited memory, low computing power, no operating system, etc. In addition, an efficient workflow for porting a neural network from e.g. Tensorflow or Keras to an embedded target is needed. Currently, several frameworks that offer such a workflow are available. This paper examines four currently available solutions and compares them: Google: tfLite, ARM: CMSIS-NN, ST: Cubex-AI, Renesas: e-Ai. The frameworks differ considerably in terms of workflow, features and performance. Depending on the application, one framework is better suited than another. Typically, neural networks that are ported to embedded platforms with these frameworks are static. This means that once they are integrated into the firmware, they can no longer be adapted. In the context of this work, possibilities of unsupervised learning on the target was investigated and is discussed in the presentation.