TRACER: task-aware risk-adaptive architecture for continual edge learning

Abstract

Current computer-vision architectures are being deployed as long-lived services, especially on edge and on-device platforms, where input distributions change with changes in environment, users, sensors, and class frequencies. In these cases, to achieve sustainable performance, continual learning is required. Also, we need to keep in mind that the process needs to be feasible under strict constraints like latency and memory. Previous experience demonstrates that device-centric measures of deployment efficiency should be used instead of proxy metrics like FLOPs, and that tail latency (e.g., p95) is a more constrained measure of deployment efficiency than mean latency. At the same time, full neural architecture search (NAS) is generally too costly to integrate into a repeated learning loop, motivating restricted, hardware-aware search strategies. This thesis presents TRACER: Task-aware Risk-adaptive Architecture for Continual Edge leaRning, a deployability-oriented continual learning pipeline that keeps a fixed feature backbone and repeatedly selects and adapts a lightweight MLP classifier head. The system follows a restricted design space with a NAS-inspired controller and a Net2Net-optimized evolutionary population so that it can adapt efficiently. The stability between tasks is ensured through risk-aware exemplar rehearsal (high-risk samples are prioritized) and knowledge distillation. Experiments on Split CIFAR-100 (10 tasks x 10 classes) and CIFAR-10 (5 tasks x 2 classes) report class-incremental (CIL) and task-incremental (Task-IL) performance. Our proof-ofconcept implementation has a final CIL mean accuracy of 0.8145 and average forgetting of 0.0341, and TIL has a final mean accuracy of 0.970 on CIFAR-10. Also, in CIFAR-100, we got 0.6136 final CIL mean accuracy, and TIL has a final mean accuracy of 0.9055 with 0.0451 forgetting. These results, which are derived from a single deterministic run, demonstrate that Lagrangian-relaxation-based constraintaware head selection, combined with risk-sensitive stabilization, provides a practical accuracy-feasibility trade-off for continual learning under explicit latency targets.