AdaSpring: Context-adaptive and Runtime-evolutionary Deep Model Compression for Mobile Applications
summary
A runtime system called AdaSpring bundles multiple compressed architectures and weights into a single retraining-free and self-evolutionary network and runs a fast runtime search strategy on-device to pick and evolve the best configuration for current context. Implementations on three platforms across five tasks show concrete gains: up to 3.1x latency reduction and 4.2x energy efficiency improvement versus hand-crafted compression, with runtime-evolution overhead ≤6.2 ms.
AI-written summary · about
related talks
29mRunning Deep Learning Models on Mobile with optimised speed
15mDesign Compact Deep Learning Models: Small is the New Big
11mHow Do We Compress Giant AI Models to Run at the Edge?
24mContext Aware Software Stacks for Mobility: Composive.ai Overview
39mCreating smaller, faster, production-worthy mobile machine learning models for Android
17mUnleash All Cores: Asymmetry-Aware Scalable DNN Inference on Mobile CPUs
30mCompactifAI: Quantum-Inspired AI Model Compression
39mAI/ML on mobile/android with TensorFlow