Generalization of winning-ticket discovery beyond image classification

Determine whether the favorable winning-ticket discovery achieved by the Improve method, which integrates iterative magnitude pruning into pool-based deep active learning, generalizes to other domains such as object detection, natural language processing, and tabular data.

Background

The paper evaluates Improve exclusively on image-classification tasks and argues that the structural correspondence between the iterative retraining loop of pool-based deep active learning and iterative magnitude pruning is task-agnostic in principle. However, the existence of winning tickets is presented as an empirical phenomenon that may depend on the architecture, dataset, task, and sparsity level. The unresolved issue is therefore whether the favorable ticket-discovery behavior observed in the evaluated image-classification settings transfers to substantially different machine-learning domains, specifically object detection, natural language processing, and tabular data.

References

Whether the favorable ticket discovery we observe generalizes to other domains (object detection, natural language processing, tabular data) remains an open question that warrants dedicated investigation.

— One Loop, Two Gains: Can Active Learning win the Lottery for Free?  (2609.10311 - Tscheschner et al., 9 Sep 2026) in Section 6, “Empirical scope and domain limitations”