Enabling Edge IoT Intelligence: A Survey on Resource-Constrained Continual Learning

Abstract

Edge intelligence is revolutionizing the Internet of Things (IoT) by migrating AI training from centralized clouds to distributed end devices. However, enabling continual learning (CL) on edge nodes faces a “rigid triangle constraint” of restricted memory, limited computation, and battery-dependent energy, heavily dictated by physical size, weight, and power (SWaP) limitations. This article systematically reviews solutions for resource-constrained CL in IoT scenarios, bridging algorithmic plasticity with these physical hardware realities. To mitigate the memory wall, we synthesize static model evolution techniques. We specifically highlight the cross-paradigm integration of parameter-efficient fine-tuning (PEFT) with low-bit quantization and structural sparsity. Furthermore, this is presented alongside knowledge distillation (KD) as a zero-buffer regularizer to effectively accommodate storage-limited microcontrollers. Crucially, we emphasize energy-efficient inference as an integral part of the CL lifecycle, as continuous perception significantly depletes the same battery budget required for future backpropagation. To sustain long-term evolution, we discuss dynamic architectures, including mixture of experts (MoE) and cloud–edge offloading, to achieve on-demand compute allocation for energy efficiency (EE). Beyond isolated nodes, addressing the bandwidth and privacy bottlenecks of IoT, we analyze an end-to-end green data lifecycle—ranging from online value filtering and physical storage compression to federated CL (FCL) in untrusted environments. Ultimately, we argue that isolated algorithmic improvements are often insufficient. Therefore, grounded by empirical hardware benchmarks, this article emphasizes compute-optimal energy frontiers and memory hierarchy integration hardware-aware neural architecture search (HA-NAS). We propose that a full-stack system–hardware codesign incorporating non-von Neumann paradigms is an essential path for realizing autonomous, sustainable evolution across diverse next-generation IoT sectors.