Deploying artificial intelligence models on low-cost edge devices requires balancing predictive accuracy with strict constraints on computational resources, such as inference latency and memory footprint. Despite growing interest in TinyML systems, limited empirical evidence exists on how these factors interact across different embedded hardware platforms. This study presents a systematic multi-objective evaluation of three lightweight AI architectures—multinomial logistic regression (MLR), multilayer perceptron (MLP), and a reduced convolutional neural network (CNN)—implemented natively on three representative platforms: ESP32-S3, Raspberry Pi Pico, and Raspberry Pi Zero W. The models were evaluated on three image classification datasets of increasing complexity (Synthetic Geometric Figures, MNIST, and Fashion-MNIST), measuring classification accuracy, inference latency, and peak memory footprint under real execution conditions. Pareto-front analysis was used to identify efficient model–platform configurations and characterize the trade-offs between predictive performance and computational resources. The results provide quantitative insight into accuracy–resource trade-offs and establish a reproducible framework for evaluating lightweight AI models on resource-constrained edge devices, supporting informed design decisions in TinyML applications.
Rojas-Carrasco et al. (Thu,) studied this question.