Router-Guided Data Selection for Efficient Deep Learning

Efficient data selection is essential for training deep neural networks when both annotation budgets and computational resources are limited. Although Active Learning (AL) reduces annotation cost, many AL pipelines introduce substantial selection overhead by relying on target-model gradients, embeddings, ensembles, or repeated stochastic passes, especially when these must be recomputed for every downstream architecture. We propose a compute efficient, target-decoupled AL framework based on a compact expert-routed selector. Given an unlabeled pool, the selector produces multiple acquisition signals in a single routed forward pass including routing uncertainty, expert disagreement, aggregate confidence, and feature-space diversity. The resulting selected indices can then be reused to train arbitrary downstream models. Experiments on various image-classification benchmarks demonstrate the effectiveness of the proposed approach compared to baseline in resource-constrained settings.