🔗 Source: arXiv

SwiftNet: Using Graph Propagation as Meta-knowledge to Search Highly Representative Neural Architectures

🚀 Technical Novelty

  • Mechanism: GRAM represents network operations as nodes in a Directed Acyclic Graph (DAG) and accumulates learned structural knowledge into a meta-graph via node-wise search and Gibbs sampling.
  • Nuance: Unlike prior cell-based NAS methods that rely on rigid blocks and width multipliers (causing accuracy drops when downsizing), GRAM enables flexible, fine-grained structure-level pruning directly on the meta-graph to remove redundant edges without retraining.

💡 Yield

  • SwiftNet achieves 2.15× higher accuracy density than MobileNet-V2 and reduces search cost by 26× compared to FBNet, reaching 63.28% top-1 ImageNet accuracy with only 53M MACs and 2.07M parameters (19.09ms latency on Pixel 1).

⚠️ Limitations

  • The meta-graph convergence relies on Gibbs sampling, which lacks a strict theoretical upper bound guarantee for the number of iterations required to converge; rigorous convergence analysis is deferred to future work.