MO-SAE:Multi-Objective Stacked Autoencoders Optimization for Edge Anomaly Detection
New optimization method slashes storage needs and power consumption for anomaly detection AI on edge devices.
A research team led by Lizhao Zhang has introduced MO-SAE (Multi-Objective Stacked Autoencoders), a novel optimization framework designed to make resource-intensive AI models viable for edge computing. The core challenge addressed is that Stacked Autoencoders (SAEs), while effective for anomaly detection in systems like industrial IoT or network security, are typically too large and power-hungry for constrained edge devices. MO-SAE tackles this by formulating the optimization as a multi-objective problem, simultaneously targeting high performance, low storage, minimal power consumption, and fast inference.
The framework employs three key techniques: model clipping to remove redundant parameters, a multi-branch exit design that allows for early inference on simpler inputs, and a matrix approximation method to compress the model's weight matrices. A multi-objective heuristic algorithm then balances these competing goals. The results are substantial: on x86 architecture, MO-SAE reduces storage and power consumption by at least 50%, improves runtime efficiency by over 28%, and achieves an 11.8% compression rate. Crucially, it maintains application-level detection performance.
Furthermore, the optimized models run efficiently on the ARM architectures common in edge hardware, showing a 15% improvement in inference speed. This makes MO-SAE a practical solution for deploying sophisticated anomaly detection in cloud-edge collaborative systems, where models in the cloud can be compressed and efficiently distributed to numerous edge nodes for real-time, local analysis.
- Achieves at least 50% reduction in storage space and power consumption on x86 hardware.
- Improves runtime efficiency by 28% on x86 and inference speed by 15% on ARM architectures.
- Uses a combination of model clipping, multi-branch exits, and matrix approximation to compress models by 11.8% without losing performance.
Why It Matters
Enables complex AI-powered anomaly detection to run locally on resource-limited devices like sensors and gateways, reducing cloud dependency and latency.