Research & Papers

Collate: New ML framework boosts edge AI accuracy by 3% under latency constraints

Heterogeneous edge devices get a 3.09% accuracy boost with zero extra training overhead.

Deep Dive

Federated learning (FL) lets multiple edge devices collaboratively train models without sharing raw data, but real-world deployments struggle with device heterogeneity—different hardware capabilities and latency requirements. Existing FL optimizations focus on training efficiency, but often ignore inference speed, leaving models too slow for time-sensitive applications like autonomous vehicles or real-time analytics.

Collate, a new training framework from a multi-institutional team, tackles this head-on. It uses a dynamic zeroizing-recovering method to automatically prune or expand each local model's architecture so it meets its device's latency budget while maximizing accuracy. A proto-corrected federated aggregation scheme then combines these heterogeneous models into one unified training process—eliminating the need for separate training runs. Benchmarks show Collate outperforms state-of-the-art methods by 1.96% average accuracy on extended models and 3.09% on shrunk models, all with negligible extra overhead. The paper appears at IEEE ICCD 2022 and will release code.

Key Points
  • Dynamic zeroizing-recovering method adapts each local model's architecture to meet individual device latency constraints while preserving accuracy.
  • Proto-corrected federated aggregation enables a single training process to handle heterogeneous models from different edge systems.
  • Achieves 1.96% and 3.09% average accuracy improvements over state-of-the-art for extended and shrunk models respectively.

Why It Matters

Enables high-accuracy AI on diverse edge devices without sacrificing latency—critical for real-time systems.

📬 Get the top 10 AI stories daily