Research map / Label-efficient and robust learning / Domain adaptation and generalization

SynEVO: A neuro-inspired spatiotemporal evolutional framework for cross-domain adaptation

ICML 2025 · Label-efficient and robust learning › Domain adaptation and generalization · approach: Mixture and distillation for domain shift · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes SynEVO, a neuro-inspired spatiotemporal network that uses curriculum-based sample reordering, an elastic common container plus task-independent personality extractor, and an adaptive coupler to enable cross-domain transfer and model evolution on urban traffic data.

Problem

Spatiotemporal learners train independent models per data source, giving limited transfer across sources and temporal shifts, and existing multi-task approaches introduce noise from unrelated tasks and cannot evolve elastically as data distributions change.

Approach

Couples curriculum learning (gradient-based ordering of sample groups) with complementary learning (elastic common container for shared patterns and contrastive personality extractor for task-specific features), gated by a distance threshold that decides whether a new domain joins the common container; dropout and weight decay are adapted from gradient differences using a neuroscience-inspired rel

Key results

• SynEVO reports best MAE on NYC (6.494 vs CMuST 6.576) and SIP (0.697 vs 0.732), and the paper claims 0.5% to 42% improvements over baselines.
• Zero-shot cross-temporal results show SynEVO beats GWN by an average of 29.8% across four datasets, per the paper.
• GPU memory on SD drops from 19533MB (CMuST) to 4252MB (SynEVO), about 21.75% of CMuST's cost as stated in the paper.

Measured effects reported in the paper

Surprising findings

• Hard-to-easy ordering (H2E) still outperforms random ordering in the reordering ablation, which the authors attribute to relations between neighboring tasks.

Limitations

• Reported gains are not uniform: SynEVO does not beat CMuST on every metric or dataset (e.g., CHI RMSE, SD MAE).
• Evaluated only on urban traffic datasets with a single backbone (GraphWaveNet).
• Key design choices (threshold kappa, p0, lambda0) are tuned per dataset via sensitivity sweeps.

Experimental setup

Datasets: NYC taxi/crowd data CHI traffic status data SIP traffic data SD traffic flow data

Benchmarks: NYC CHI SIP SD

Metrics: MAE RMSE MAPE

Baselines: STGCN STGODE GWN STTN AGCRN ASTGCN CMuST

Builds on: GraphWaveNet curriculum learning contrastive learning Information Bottleneck

Closest papers by meaning

• Socialized Coevolution: Advancing a Better World through Cross-Task Collaboration (ICML 2025 · similarity 0.85)
• Reinforced Cross-Domain Knowledge Distillation on Time Series Data (NeurIPS 2024 · similarity 0.84)
• UniSTD: Towards Unified Spatio-Temporal Learning across Diverse Disciplines (CVPR 2025 · similarity 0.84)
• Learning the Plasticity: Plasticity-Driven Learning Framework in Spiking Neural Networks (NeurIPS 2025 · similarity 0.84)
• SEEA-R1: Tree-Structured Reinforcement Fine-Tuning for Self-Evolving Embodied Agents (NeurIPS 2025 · similarity 0.84)
• Get Rid of Isolation: A Continuous Multi-task Spatio-Temporal Learning Framework (NeurIPS 2024 · similarity 0.83)