NuWa: Deriving Lightweight Class-Specific Vision Transformers for Edge Devices

Researchers introduce NuWa, a novel model compression technique that derives lightweight, class-specific Vision Transformers optimized for edge devices. By identifying and removing class-detrimental weights through self-knowledge purification, NuWa achieves up to 29% accuracy improvements on specialized tasks while reducing pruning costs by 99.83% compared to existing methods.

NuWa addresses a practical bottleneck in deploying AI models to edge devices like drones and autonomous vehicles. Current compression methods treat all classes equally, leaving specialized devices burdened with irrelevant knowledge. The research identifies a critical insight: certain weights actively harm performance for specific classes, and removing them paradoxically improves accuracy. This discovery challenges conventional assumptions about neural network pruning. The technical contribution extends beyond pruning—NuWa's closed-form optimization eliminates expensive iterative retraining, achieving 33.69x speedup compared to state-of-the-art methods. For deployment scenarios where edge devices handle narrow tasks (e.g., detecting specific vehicle types or pedestrians), this efficiency gain translates directly to reduced computational overhead, lower power consumption, and faster response times. The 0.61% average accuracy loss against training-dependent methods represents acceptable trade-off for 99.83% cost reduction. Industry impact centers on making AI deployment economically viable for resource-constrained environments. Current compression workflows consume significant compute resources, limiting adoption for smaller organizations and research teams. NuWa democratizes this capability by reducing barriers to deployment. The method particularly benefits autonomous systems and IoT applications where specialized models serve specific functions. Looking ahead, this work signals growing interest in task-specific rather than general-purpose model compression, reflecting real-world deployment constraints. The open-source release enables rapid adoption and validation across different edge hardware platforms.

• →NuWa achieves 29% accuracy improvement on class-specific vision tasks by removing weights that harm specialization
• →Closed-form optimization reduces pruning costs by 99.83% and achieves 33.69x speedup versus existing training-dependent methods
• →The technique requires no post-pruning retraining, making lightweight model deployment economically practical for edge devices
• →Class-detrimental weights discovery challenges conventional pruning assumptions and opens new optimization research directions
• →Open-source availability enables rapid adoption for autonomous systems, drones, and IoT applications with narrow task requirements