A Survey of Model Compression Techniques for TinyML Applications

Abstract

The convergence of embedded systems and artificial intelligence has led to the rapid growth of TinyML, which enables ML inference on highly resource-constrained devices. However, deploying modern DL models in these environments poses signif-icant challenges due to limited memory, processing power, and energy availability. Model compression techniques have emerged as essential tools to bridge this gap, allowing for the deployment of efficient, low-latency, and accurate models at the edge. This survey provides a comprehensive and structured review of key model compression strategies-including low-rank factorization, neural architecture search, pruning, quantization, knowledge distillation, and other emerging methods-highlighting their principles, mathematical formulations, and application relevance to TinyML. We also discuss the trade-offs and challenges as-sociated with these methods, such as accuracy loss, hardware-software co-design, and deployment constraints. Our goal is to equip researchers and practitioners with a practical reference that informs the design of lightweight AI solutions suitable for the next generation of intelligent edge devices.