PredRNN: A Recurrent Neural Network for Spatiotemporal Predictive Learning

arXiv 2021 32 method

TLDR

PredRNN introduces a recurrent network with decoupled memory cells and zigzag memory flow for spatiotemporal predictive learning, achieving competitive results on five datasets.

Reasoning

The paper presents a novel architecture with explicit memory decoupling and cross-layer communication, supported by ablation studies and strong empirical results on multiple benchmarks. However, it does not connect to broader world model concepts or real-world deployment, and the contribution is limited to predictive learning without interactive or reinforcement learning contexts.

Read-first score

Read-first score 32, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 10.

Methodology quality 25%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=ablation,dataset,result

Recency 8%

Uses a gentle age decay so recent papers surface without erasing older foundations. 2021

Reproducibility 25%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=dataset

Topical relevance 42%

14.3

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Field roles

Candidate

Rank sensitivity

Stability: volatile; rank range: 56.

Keyword Scores

world dynamics prediction

video world model

world model

generative world model

world simulator

interactive world model

model-based reinforcement learning world model

Resources

Paper