A survey paper on Dynamic Neural Networks in Computer Vision, providing a taxonomy based on network adaptivity types: output, computation graph, and input.
https://arxiv.org/abs/2501.07451
🔍 Methods explored this Paper:
→ Early Exits networks allow intermediate predictions for "easy" inputs, reducing unnecessary computations.
→ Dynamic Routing enables conditional computation paths through techniques like Mixture-of-Experts.
→ Token Skimming focuses on efficient processing by dynamically reducing redundant tokens in Vision Transformers.
→ The research emphasizes applications in Multi-modal Sensor Fusion, highlighting benefits like sample-aware efficiency and noise robustness.
-----
💡 Key Insights:
→ Different inputs require varying computational resources - simple images can be processed with fewer computations
→ Dynamic Neural Networks can adapt their architecture on-the-fly based on input complexity
→ Sensor Fusion particularly benefits from dynamic approaches due to varying sensor reliability
Share this post