RAMP: Robust Adaptive Mixed-Precision Quantization for Edge CPU Vision Models
RAMP finds Jensen-Shannon Divergence plus clustering to be the most reliable path for INT8 layer selection on edge CPUs.
The paper tests 13 layer-sensitivity metrics across four vision networks and validates policies on two ARM64 platforms. Gradient-based methods fail in 4 of 8 model-hardware setups, while weight statistics fail in 2; Jensen-Shannon Divergence has no catastrophic failures in their study. The authors pair it with K-Means instead of fixed thresholds, reporting near-lossless accuracy and a mean 1.81x speed-up over full precision. They also warn that skipping layers with small individual speed-ups can hurt by fragmenting the graph and breaking operator fusion. ArXiv · AI/CL/LG's note
The paper tests 13 layer-sensitivity metrics across four vision networks and validates policies on two ARM64 platforms. Gradient-based methods fail in 4 of 8 model-hardware setups, while weight statistics fail in 2; Jensen-Shannon Divergence has no catastrophic failures in their study. The authors pair it with K-Means instead of fixed thresholds, reporting near-lossless accuracy and a mean 1.81x speed-up over full precision. They also warn that skipping layers with small individual speed-ups can hurt by fragmenting the graph and breaking operator fusion. ArXiv · AI/CL/LG's note
score 4