Meta-Learning Enhanced Deep Neural Networks for Cross-Domain Generalization in Dynamic and Non-Stationary Environments

Abstract

Deep neural networks deployed in intelligent systems commonly encounter domain shifts that are neither independent nor stationary: sensor characteristics change, class priors evolve, decision boundaries drift, and previously observed regimes recur. While conventional domain generalization optimizes a fixed predictor prior to deployment, online adaptation typically does not have an initialization or control policy for cross-domain transfer. In this article, we present a novel adaptive neural architecture with delayed supervision, called MetaDyG-Net, which is enhanced by meta-learning. The method learns (i) an encoder initialization that can be transferred across tasks and (ii) a small drift-aware controller which maps the discrepancy between the features moments, the predictive entropy, and the uncertainty to the layer-wise adaptation rates and the replay strength. Deployment is the fusion of a reservoir memory, Fisher-weighted elastic anchoring, class-prototype fusion and uncertainty-adaptive probability tempering. A first-order episodic objective is used to let the model see pseudo-unseen source domains while explicitly learning the stability-plasticity trade-off. Evaluation is performed on five independent seeds on two controlled cross-domain streams: DS-Digits, which is created from real handwritten-digit images with abrupt, gradual, recurrent and compound visual shifts, and DMS-24, a nonlinear multi-sensor stream with covariate, prior and concept drift. Under test-then-train evaluation, MetaDyG-Net attains 67.09 +/- 0.90% accuracy on DS-Digits and 62.20 +/- 0.60% on DMS-24. The gains are 0.61 and 0.65 percentage points relative to the strongest fixed-replay baseline, respectively, and the DMS-24 gain is significant after Holm correction, whereas the DS-Digits gain, relative to the MLDG+Replay baseline, is directionally consistent but not significant in a two-sided paired t-test. More importantly, the expected calibration error drops to 8.41% and 4.23% and the worst-regime accuracy increases to 35.13% and 51.42%. The ablation results show that replay is the most important contributor, and that meta-control, Fisher anchoring, prototype fusion and selective extra updates contribute less and less. The protocol-aware synthesis of SWAD, Fishr, DRAIN, CoTTA, and Wild-Time distinguishes between static DG, future-domain extrapolation, continual test-time adaptation without labels, and delayed-label adaptation; published scores are only used as within-study comparisons, and not as a common leaderboard, due to the different benchmarks and supervision regimes. These results validate drift-aware meta-control as a viable computational-intelligence mechanism for developing adaptive neural systems that are accurate, calibrated and auditable when operating in a non-stationary cross-domain setting.