Energy-Aware TinyML for Ambient-Powered, Hardware-Constrained IoT Nodes

Abstract

Resource-constrained Internet of Things (IoT) devices are increasingly expected to perform intelligent sensing under strict limitations in memory, computation, and energy. For image-based applications, transmitting raw images can consume substantially more communication energy and time than transmitting a compact latent representation. This thesis investigates lightweight convolutional autoencoders for on-device image compression on microcontroller-class hardware. A design space of 70 convolutional autoencoder models was evaluated on the MNIST dataset. The models differed in encoder filter configuration, bottleneck dimension, and compression ratio. Reconstruction quality and task utility were measured by normalized mean squared error (NMSE) and downstream classifier accuracy. The baseline models were further optimized using post-training quantization (PTQ), quantization-aware training (QAT), and a two-stage magnitude-based pruning method. Different optimization orders were compared with respect to model size, reconstruction quality, classifier accuracy, and estimated energy. An analytical energy model was used to estimate microcontroller active energy associated with encoder computation and payload transfer under explicitly stated assumptions. Pareto-front analysis identified non-dominated candidate models offering different tradeoffs between reconstruction quality and estimated energy. A selected INT8 encoder was deployed on an Ambiq Apollo3 microcontroller using TensorFlow Lite for Microcontrollers and the NeuralSPOT software stack. Hardware experiments confirmed successful encoder inference and provided measured inference-time results. The results show that bottleneck size and the choice of optimization method strongly influence the trade-off between reconstruction quality, payload size, and deployment cost. Quantization reduced numerical precision and memory footprint, while unstructured pruning introduced sparsity but did not necessarily reduce execution time on dense microcontroller kernels. The study therefore carefully distinguishes analytical energy estimates from measured hardware results. This work provides a reproducible framework for comparing learned compression models on resource-constrained embedded devices. The evaluation is limited to MNIST data, analytical communication-energy assumptions, host-side serial data transfer, and a single microcontroller platform. A complete autonomous energy-harvesting and wireless IoT