Sigma-Branch: Hierarchical Single-Path Network Reconstruction for Dynamic Inference with Reduced Active Parameters

Researchers introduce Sigma-Branch, a neural network restructuring framework that reduces per-inference active parameters by 58-60% while maintaining full model capacity in memory. The approach uses hierarchical routing and binary tree architecture to enable efficient edge deployment without permanent model compression trade-offs.

Sigma-Branch addresses a critical bottleneck in edge AI deployment: the cost of transferring dense network weights from off-chip memory during inference. Traditional compression techniques reduce model size permanently, sacrificing capacity for efficiency. This research decouples those constraints by keeping the complete model in storage while activating only a single computational path per inference through hierarchical routing. The technical innovation uses spherical k-means clustering to initialize a binary tree structure where inputs follow routed paths to specialized leaf nodes, balancing computational efficiency with model expressiveness. The framework demonstrates consistent results across diverse architectures—ResNet-50 on vision tasks and PointNet++ on 3D point clouds—suggesting domain-agnostic applicability. The 14-23 percentage point improvement in active-parameter reduction compared to static pruning methods indicates a meaningful algorithmic advance. This matters for edge computing, IoT devices, and resource-constrained environments where memory bandwidth rather than computational throughput limits performance. The ability to maintain full model capacity while reducing inference footprint could accelerate AI deployment in autonomous systems, mobile devices, and embedded applications. However, the approach introduces router overhead and added complexity during fine-tuning, which may limit adoption in resource-scarce scenarios. Future work should examine how hierarchical routing performs under latency constraints and whether the method scales effectively to transformer architectures gaining prominence in modern AI systems.

• →Sigma-Branch reduces active inference parameters by 58-60% while preserving full model capacity, outperforming traditional pruning methods by 14-23 percentage points
• →The framework uses hierarchical binary tree routing with specialized leaf nodes, enabling single-path inference execution
• →Spherical k-means clustering jointly initializes router weights and channel allocations, streamlining the restructuring process
• →Cross-domain validation on vision and 3D point-cloud tasks demonstrates framework generalization beyond single-architecture evaluation
• →Approach decouples memory traffic from total parameter count, addressing fundamental edge deployment constraints