💻 computer science

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

Dyna-Pruner is an end-to-end framework that dynamically co-prunes redundant input regions and corresponding model structures based on input-dependent importance, achieving up to 70% FLOPs reduction and a 2.5× inference speedup on edge devices with negligible accuracy loss for spatio-temporal media prediction tasks.

Original authors: Fuyan Zhang, Yuqi Li, Yingli Tian, Edmond S. L. Ho

Published 2026-06-16

📖 4 min read☕ Coffee break read

CC0 1.0

Original authors: Fuyan Zhang, Yuqi Li, Yingli Tian, Edmond S. L. Ho

Original paper dedicated to the public domain under CC0 1.0 (http://creativecommons.org/publicdomain/zero/1.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a weather forecaster looking at a massive, high-definition map of the entire Earth. Your job is to predict what the weather will look like in the next few hours.

The Problem: The "One-Size-Fits-All" Approach
Currently, most computer models act like a very diligent but inefficient student. They look at the entire map, pixel by pixel, with the same amount of brainpower, regardless of what's happening.

They stare intensely at a calm, clear blue sky over the ocean (where nothing is changing).
They stare just as hard at a chaotic, swirling storm cell (where everything is changing).

This is a waste of energy. The computer is doing heavy math on the "boring" parts of the map, leaving it too slow and expensive to run on small devices like drones, satellites, or city traffic sensors.

The Solution: Dyna-Pruner
The paper introduces a new system called Dyna-Pruner. Think of this as a smart "spotlight" and a "switchboard" working together.

Instead of treating the whole map equally, Dyna-Pruner does two things simultaneously for every single image it sees:

The Data Spotlight (Data Mask): It looks at the input map and asks, "Where is the action?" It shines a bright light on the storm clouds and traffic jams, while dimming the lights on the calm ocean and empty streets. It essentially tells the computer, "Ignore the quiet parts; they don't matter right now."
The Switchboard (Model Mask): This is the clever part. Usually, if you tell a computer to ignore part of an image, it still uses all its internal gears to process it. Dyna-Pruner goes further. It physically turns off the specific gears (computational units) that would have been working on those quiet parts.

The Creative Analogy: The Restaurant Kitchen
Imagine a busy restaurant kitchen (the computer model) trying to cook meals for thousands of customers (the data).

The Old Way: The head chef forces every single cook to chop vegetables, even if the customer only ordered a salad. The cooks are busy, but much of the work is wasted on customers who didn't order it.
The Dyna-Pruner Way:
- The Spotlight: The waiter (Data Mask) looks at the order and says, "Table 5 just wants a salad; Table 6 wants a steak."
- The Switchboard: The head chef (Model Mask) immediately sends the salad chefs home for the break and tells the steak chefs to focus only on the steak.
- The Result: The kitchen runs 2.5 times faster, uses less electricity, and the food still comes out perfect because the right people were working on the right tasks.

How It Works (The Magic Trick)
The paper explains that this isn't just a rulebook; the computer learns how to do this.

It uses a "Shared Importance" system. The computer learns a map of "importance" (where the interesting stuff is).
It uses this same map to decide both what to look at (the data) and which tools to use (the model).
Because these two decisions are linked, the system doesn't accidentally turn off a tool that it actually needs. They work in sync.

The Results
The researchers tested this on three different types of problems:

Weather: Predicting storms and rain.
Traffic: Predicting city congestion.
General Media: Various spatio-temporal data.

They tried it on three different types of computer brains (CNNs, RNNs, and Transformers). The results were impressive:

Speed: It ran 2.5 times faster on small, edge devices (like the NVIDIA Jetson AGX Orin).
Efficiency: It reduced the amount of math needed by up to 70%.
Accuracy: Despite doing so much less work, it barely lost any accuracy (less than 1% error increase). It was almost as good as the heavy, slow version.

Why It Matters
The paper concludes that by letting the model "choose" where to focus its energy based on the specific situation, we can run powerful prediction tools on small, battery-powered devices. This makes real-time weather forecasting and traffic monitoring possible in places where heavy computers can't go.

In short: Dyna-Pruner teaches computers to stop wasting energy on the boring stuff and focus all their power on the action.

Technical Summary: DYNA-PRUNER

Problem Statement
Spatio-temporal prediction is critical for applications such as weather nowcasting and urban traffic monitoring. However, modern deep learning models (CNNs, RNNs, Transformers) face a fundamental inefficiency: they employ a "dense" computation paradigm that processes all input pixels uniformly. This approach ignores the inherent redundancy in real-world spatio-temporal data, where vast regions (e.g., calm seas or clear skies) contain negligible dynamic information compared to localized active regions (e.g., storm cells). Consequently, these models waste significant computational resources and energy on low-information areas, creating a bottleneck for real-time deployment on resource-constrained edge devices. Existing compression techniques, such as static weight pruning or standalone data simplification, fail to dynamically link input information density with the model's active computational path.

Methodology
The authors propose Dyna-Pruner, an end-to-end framework for input-adaptive data–model co-pruning. Unlike static methods, Dyna-Pruner dynamically generates sparse sub-networks tailored to the specific information density of each input sample.

Core Mechanism: The framework utilizes a Shared-Importance Synchronization Mechanism. A lightweight Mask Generator produces a continuous Importance Score Map ( $S$ ) from the input data. This map serves as a unified bridge for two synchronized actions:
1. Data Pruning: A Data Mask ( $M_{data}$ ) is derived from $S$ to nullify redundant spatial regions in the input.
2. Model Pruning: A Model Mask ( $M_{weight}$ ) is generated by aggregating importance scores within the receptive fields of structural units (e.g., convolutional filters). Units with low aggregated importance are pruned.
Training Strategy: To handle the non-differentiable nature of discrete masking, the framework employs a continuous proxy variable ( $S$ ) during training, using the Straight-Through Estimator (STE) to simulate binarization for inference.
Optimization: The system is trained via a unified objective function that balances task performance (e.g., MSE) with sparsity constraints. The loss function includes the task loss plus L1-norm regularization terms for both the importance score map and the unit importance scores, controlled by hyperparameters $\lambda_d$ and $\lambda_w$ .

Key Contributions

Unified Framework: The paper introduces the first dynamic collaborative pruning framework that unifies data sparsification and model structural compression into a single end-to-end optimizable objective, moving beyond traditional input-agnostic compression.
Synchronization Mechanism: It proposes a novel shared-importance synchronization mechanism that enables per-input adaptation through receptive-field aggregation and STE-based training, ensuring the model focuses computational budget exclusively on high-value, physically active regions.
Empirical Validation: Extensive evaluations demonstrate the framework's ability to significantly reduce computational complexity while maintaining high prediction accuracy across diverse architectures.

Experimental Results
The framework was evaluated on three benchmarks: WeatherBench (medium-range weather), SEVIR (nowcasting), and TaxiBJ (traffic). It was integrated with three distinct backbone architectures: SimVP (CNN), ConvLSTM (RNN), and TAU (Transformer).

Efficiency: Dyna-Pruner achieved up to a 70% reduction in FLOPs and a 2.5× speedup in inference latency on NVIDIA Jetson AGX Orin hardware.
Accuracy: The method maintained negligible accuracy loss, with prediction errors increasing by less than 1–2% compared to dense models.
Comparison: Compared to standard Model Pruning (MP) baselines at similar sparsity levels, Dyna-Pruner significantly outperformed them. While MP incurred accuracy degradations of 4–7%, Dyna-Pruner kept degradation within 1–2%.
Robustness: The framework demonstrated resilience against Gaussian noise and simulated mask failures, outperforming random pruning baselines in noisy conditions.
Generality: The approach proved effective across CNN, RNN, and Transformer backbones, confirming its model-agnostic nature.

Significance and Claims
The paper claims that Dyna-Pruner addresses the critical mismatch between dense computation and the sparse, dynamic nature of spatio-temporal data. By establishing a dynamic link between input information content and the active computational path, the framework enables automated, resource-aware architecture optimization. The authors assert that the synergy between data and model streamlining is more effective than isolated compression techniques. This work provides a practical paradigm for deploying high-fidelity predictive models on edge devices and IoT sensors, where real-time insights are essential but computational resources are limited.

Technical Summary: DYNA-PRUNER

More like this