Improving integration of ML models in embedded systems

Sammanfattning

The rise of machine learning and artificial intelligence has led to increased demands on using different platforms to develop and run machine learning models. This has led to an increasing interest in the field of Machine Learning Operations (MLOps) which aims to ensure a more modular approach to the machine learning pipeline where for example the machine learning model can be updated while it is running live. Machine learning models can be saved in several formats depending on the framework that was used to construct them. Thus this leads to a problem regarding portability of machine learning models due to incompatible formats. One solution to this problem is to use the Open Neural Network Exchange (ONNX) format which is widely accepted by most machine learning frameworks. However, problems still persist due to the incompatibility for embedded systems to run ONNX models. Thus this thesis aims to solve the problem of model portability on embedded systems by providing a method to update models running on embedded systems by extracting the necessary information from an ONNX model and packing it into a data exchange format such that it can be transferred to the embedded system and the model can be updated using the extracted data. From the results it is seen that for simple feedforward neural network models this method to update the models is successful and the time taken to update the model increases linearly with an increase in the size of the models being updated.