DYNA-PRUNER: Input-Adaptive Data-Model Co-Pruning for Efficient and Scalable Spatio-Temporal Media Prediction

A new framework called Dyna-Pruner can automatically strip redundant data and model components from spatio-temporal prediction systems, cutting computational load by up to 70% while preserving accuracy, according to research posted to arXiv [1]. The work targets a persistent bottleneck in real-time media analysis: dense neural networks waste cycles on unchanging inputs such as clear skies or calm seas. The authors propose an end-to-end method that generates coupled masks to prune both irrelevant input regions and their associated computational units at inference time [1]. The technique integrates with convolutional, recurrent, and Transformer backbones, producing per-sample sparse sub-networks without manual architecture tuning [1]. On the WeatherBench, SEVIR, and TaxiBJ benchmarks, Dyna-Pruner reduced floating-point operations by up to 70% and delivered a 2.5× speedup on an NVIDIA Jetson AGX Orin edge device. Accuracy degradation stayed below 1% across tested configurations [1]. The paper was submitted to arXiv’s computer-vision section on 13 June 2026 [1]. arXiv, which began operating in August 1991, hosts preprints across physics, computer science, and related fields and does not conduct peer review [6]. The repository passed two million articles by the end of 2021 and currently receives roughly 24,000 submissions per month [6]. The Dyna-Pruner manuscript appears alongside community-built tools such as Bibliographic Explorer and CORE Recommender, which are delivered through the arXivLabs framework that allows third-party developers to add features to article pages [4][5]. Edge deployment of spatio-temporal models has drawn attention as weather nowcasting and traffic monitoring demand low-latency inference on hardware with tight power budgets. The Dyna-Pruner authors argue that input-dependent redundancy is the core mismatch between dense computation and real-time requirements, and their shared-importance synchronization mechanism addresses it by jointly deciding which data and which filters to retain for each sample [1].