Multi-Domain Learning with Global Expert Mapping

2026-04-22 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision · Depth: Expert, extended

Summary

The paper introduces Global Expert Mapping (GEM), a planner-compiler framework designed to enhance multi-domain object detection (MDOD) by optimizing Mixture-of-Experts (MoE) routing. Traditional MoE models struggle with diverse datasets due to load-balancing mechanisms that conflict with expert specialization, leading to redundant representations and poor performance on underrepresented domains. GEM addresses this by replacing the learned router with a global scheduler, using linear programming relaxation (LPR) to compute a fractional assignment of datasets to experts, followed by hierarchical rounding to convert this into a deterministic, capacity-aware mapping. Integrated into the DINO object detector, GEM-DINO achieves state-of-the-art performance on the UODB benchmark, outperforming existing MoE baselines by +2.1 AP over SoftMoE, +0.8 AP over REMoE, and +1.2 AP over MoE++. It also demonstrates superior parameter efficiency and effectively resolves task interference in few-shot adaptation scenarios.

Key takeaway

For research scientists developing robust computer vision models for diverse, real-world applications, you should consider adopting GEM's planner-compiler approach. This method directly addresses the limitations of traditional MoE load-balancing, enabling true expert specialization and significantly improving performance on challenging multi-domain and few-shot tasks. Implementing GEM can lead to more stable training, better interpretability, and superior accuracy compared to existing MoE routing strategies.

Key insights

GEM's planner-compiler framework optimizes MoE routing for multi-domain learning, achieving superior specialization and performance.

Principles

• Global scheduling outperforms local routing for MoE specialization.
• Capacity constraints prevent expert collapse and task interference.
• Deterministic assignment improves interpretability and stability.

Method

GEM uses linear programming relaxation (LPR) to plan fractional dataset-to-expert assignments, then applies hierarchical rounding to convert this into a deterministic, capacity-aware mapping, eliminating load-balancing losses.

In practice

• Integrate GEM into DINO for multi-domain object detection.
• Use a frozen backbone to compute dataset-expert affinity scores.
• Apply class-level specialization for fine-grained tasks.

Topics

• Multi-Domain Learning
• Mixture-of-Experts
• Global Expert Mapping
• Linear Programming Relaxation
• Hierarchical Rounding

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• VISION MoE Routing Explained in 5 Sentences — MoE models exhibit a "routing distraction" where visual inputs fail to activate relevant reasoning experts, leading to performance degradation.
• ProbMoE: Differentiable Probabilistic Routing for Mixture-of-Experts — ProbMoE offers a differentiable probabilistic routing solution for MoE models, improving expert utilization and performance.
• Instance-Aware Knowledge Distillation for Semi-Supervised Learning of an On-Board Multi-Task Dense Prediction Model for Collision Avoidance System — Instance-aware knowledge distillation improves lightweight models for edge-deployed multi-task dense prediction by mitigating teacher bias.
• STAR: Rethinking MoE Routing as Structure-Aware Subspace Learning — STAR rethinks MoE routing as structure-aware subspace learning for stable expert specialization.
• RaMP: Runtime-Aware Megakernel Polymorphism for Mixture-of-Experts — Optimizing MoE inference requires dynamic kernel configuration based on runtime expert routing, not just batch size.
• Toward Calibrated Mixture-of-Experts Under Distribution Shift — Expert calibration ensures hard-routed MoE calibration under shift, but soft-routed MoE requires adversarial reweighting for calibration.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.