Enhanced Vector Quantization for Embedded Machine Learning: A Post-Training Approach With Incremental Clustering

TinyML enables the deployment of Machine Learning (ML) models on resource-constrained devices, addressing a growing need for efficient, low-power AI solutions. However, significant challenges remain due to strict memory, processing, and energy limitations. This study introduces a novel method to optimize Post-Training Quantization (PTQ), a widely used technique for reducing model size, by integrating Vector Quantization (VQ) with incremental clustering. While VQ is a technique that reduces model size by grouping similar parameters, incremental clustering, implemented via the AutoCloud K-Fixed algorithm, preserves accuracy during compression. This combined approach was validated on an automotive dataset predicting CO2 emissions from vehicle sensor measurements such as mass air flow, intake pressure, temperature, and speed. The model was quantized and deployed on Macchina A0 hardware, demonstrating over 90% compression with negligible accuracy loss. Results show improved performance and deployment efficiency, showcasing the potential of this combined technique for real-world embedded applications.