(POSTER) Insights from Executing TinyML Models on Smartphones and Microcontrollers
Harman M. Singh, Shrishailya Agashe, Shreyans Jain, Surjya Ghosh, Aditya Challa, Sravan Danda, Sougata Sen · 2023
In this paper, we empirically compare the system-level performances of executing a machine learning task on a resource-constrained IoT device and compare its performance to offloading the task to a more capable device, a smartphone. Our results indicate that although the resource-constrained device cannot run complex machine learning models, they can provide reasonable accuracy using similar models that can load on their memory. For running simpler models on the microcontroller, the best case accuracy for the machine learning task was 94.32% (SD: 1.7%). Furthermore, local computation resulted in almost 50% lesser current draw as compared to the offloading. These observations make a case for adopting an adaptive approach to ensure that applications meet the energy-accuracy-latency balance.