DYNA-PRUNER: Input-Adaptive Data-Model Co-Pruning for Efficient and Scalable Spatio-Temporal Media Prediction
Researchers have developed Dyna-Pruner, a novel framework designed to optimize spatio-temporal prediction models for efficiency and scalability. This system adaptively prunes both data and model structures based on input characteristics, creating sample-specific sparse sub-networks. Dyna-Pruner has demonstrated significant reductions in computational load, achieving up to a 70% decrease in FLOPs and a 2.5x speedup on hardware like the NVIDIA Jetson AGX Orin, with minimal impact on accuracy. AI
IMPACT This research could enable more efficient real-time deployment of complex AI models for tasks like weather forecasting and traffic monitoring.