Uncertainty-Aware Adaptive Routing for Long-Tailed Recognition (UARL)

Abstract

Long-tailed visual recognition remains an open challenge: models trained under empirical risk minimization collapse toward majority classes, and existing mixture-of-experts (MoE) remedies assign experts to fixed network depths regardless of per-sample difficulty or per-class frequency. We identify three neglected gaps: (i) rigid, data-agnostic depth allocation, (ii) weak element-wise fusion operators that cannot selectively recombine cross-depth semantics, and (iii) the absence of uncertainty quantification in the distillation pipeline, which permits confidently wrong teachers to propagate misleading gradients to tail-class examples. To address these gaps, we propose Uncertainty-Aware Adaptive Routing for Long-Tailed Recognition (UARL), built on three tightly coupled components. Adaptive Sparse Routing (ASR) learns a lightweight gating network that routes each sample to a dynamically chosen subset of depth heterogeneous experts via a differentiable top-k sparse gate. Cross-Depth Attention Fusion (CDAF) replaces Hadamard multiplication with a scaled cross-attention module that queries deep semantics with shallow geometry, producing richer fused representations. Epistemic-Calibrated Distillation (ECD) weights the knowledge distillation loss by the epistemic uncertainty of each teacher, estimated via stochastic Monte Carlo dropout, so uncertain teachers contribute proportionally less signal. The proposed method achieves high performance on benchmark datasets.