Answers are grounded in this paper and cite their sources.
Multi-task deep learning improves generalization by sharing information across related tasks, but common sharing mechanisms often encode task relationships implicitly, which can obscure negative transfer and limit interpretability. This paper proposes ALE–Fréchet, an explainability-guided adaptive sharing framework for multi-task regression. The method estimates task relatedness from Accumulated Local Effects (ALE) profiles and uses the Fréchet distance between these profiles to regulate soft parameter sharing during training. Tasks with similar feature-effect patterns are encouraged to share information, whereas dissimilar tasks remain more independent. The learned similarity structure is directly inspectable through ALE curves and task-similarity matrices, providing a transparent alternative to latent gating or expert-routing mechanisms. ALE–Fréchet is evaluated on six benchmark settings, including two synthetic datasets and real-world electricity, exchange-rate, ATM-withdrawal, and traffic forecasting scenarios, against single-task MLPs, hard and soft sharing, Cross-stitch networks, MMoE, PLE, and MTAN. Across these benchmarks, the validation-selected ALE–Fréchet configuration family has the lowest observed average RMSE and MAE ranks; this includes an uncoupled Polynomial configuration with λ = 0 and similarity updates disabled, and does not establish a first-place ranking for active coupling ( λ > 0 ). Friedman tests indicate overall differences among methods, but no reported pairwise comparison is significant after Holm correction (minimum p Holm = 0.094 ). The method is therefore best understood as a competitive and interpretable alternative when task transfer must be inspected, albeit at the cost of higher computational overhead. Ablation results show that less frequent similarity recomputation can reduce this cost. Overall, the results support explanation-derived functional similarity as a useful inductive bias for transparent knowledge sharing in heterogeneous multi-task learning.
No takes yet. Share an insight, caveat, or question.
Hidalgo et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: