Abstract

The paper examines TinyML, an approach that runs machine-learning inference directly on low-power microcontrollers and edge Internet of Things devices. The benefits of local inference include lower dependence on permanent connectivity, reduced cloud data transfer, lower latency for simple reactive scenarios, and greater control over raw sensor data. The main constraints are systematized: memory, compute capacity, energy budget, and hardware heterogeneity. A practical development cycle is proposed, covering task selection, data collection, model compression and quantization, memory profiling, latency measurement, and field validation.