A Survey on Dynamic Neural Networks: from Computer Vision to Multi-modal Sensor Fusion

2025-01-13

AI Summary

Conventional vision model compression methods struggle to accommodate varying image complexity and fail to address key challenges in multi-sensor fusion—namely environmental adaptability, noise robustness, and prioritized information processing. Method: This paper presents a systematic survey of dynamic neural networks, proposing a unified taxonomy organized along three adaptive dimensions: output, computation graph, and input. It establishes the first comprehensive logical taxonomy spanning computer vision to multimodal sensor fusion. Through integrated analysis of mechanisms—including early exiting, skip connections, conditional computation, sparse activation, gating subnetworks, and dynamic cross-modal weight fusion—the work synthesizes over 100 studies into a structured knowledge graph. Contribution/Results: The survey identifies consistent empirical patterns: dynamic networks reduce computational cost by 30–70% in edge deployment. It formalizes new paradigms—noise suppression, information scheduling, and cross-modal adaptation—and advances sensor fusion toward input-aware, dynamically reconfigurable architectures.

Technology Category

• Computer Vision: Multi-modal Vision
• Intelligent Robots: Multimodal Perception & Sensor Fusion
• Machine Learning: Learning on the Edge & Model Compression

Application Category

• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Systems and Infrastructure for Web, Mobile and WoT: Applied ML and AI for Web-based mobile applications

Abstract

Model compression is essential in the deployment of large Computer Vision models on embedded devices. However, static optimization techniques (e.g. pruning, quantization, etc.) neglect the fact that different inputs have different complexities, thus requiring different amount of computations. Dynamic Neural Networks allow to condition the number of computations to the specific input. The current literature on the topic is very extensive and fragmented. We present a comprehensive survey that synthesizes and unifies existing Dynamic Neural Networks research in the context of Computer Vision. Additionally, we provide a logical taxonomy based on which component of the network is adaptive: the output, the computation graph or the input. Furthermore, we argue that Dynamic Neural Networks are particularly beneficial in the context of Sensor Fusion for better adaptivity, noise reduction and information prioritization. We present preliminary works in this direction.

Problem

• Model Compression
• Computational Variance
• Multi-Sensor Integration

Innovation

• Dynamic Neural Networks
• Model Compression
• Multi-Sensor Fusion