A Self-Evolving Neural Network Architecture for Autonomous Decision-Making in Low-Power Edge Devices

Conventional neural networks are structurally frozen for the duration of their operation, are trained once, and are deployed once. These static models need to be externally retrained when the deployment environment changes, such as when the distribution of sensor input drifts, the environment changes, or user behavior patterns change. For real-world edge devices, where large datasets, high-power processors, and constant internet access are not available, this restriction becomes crucial. A Self-Evolving Neural Network (SENN) architecture that automatically modifies its internal structure while being deployed on low-power edge devices is proposed in this paper. Without the need for external supervision or cloud-dependent retraining, SENN replaces subpar neural networks with periodic mutations and evaluations using on-device neuroevolution. Hardwareaware constraints are incorporated into the evolutionary process to guarantee that no evolved model surpasses the underlying microcontroller's memory, latency, or power limitations. The following are the contributions made by this work: (1) An inference engine with a fully autonomous neural evolution pipeline built right in. (2) Evolutionary model restructuring without retraining datasets is made possible by a resource-bounded mutation and fitness scoring procedure. (3) A deployment-oriented architecture that supports edge processors with integer-only execution, less than 1 MB of RAM, and a power budget of less than 100 mW. In comparison to static inference models, SENN offers long-term adaptability, eliminates retraining costs, and uses 40–60% less energy. For mission-critical edge environments like wearable monitoring, autonomous infrastructure, disaster-resilient sensor networks, and smart agriculture systems, this study develops a framework for continuously self-optimizing embedded intelligence.