DAQ: Decision-Aligned Post-Training Quantization for Spiking Transformers
DAQ cuts reported energy use by 66.2% with 4-bit weights and membrane variables in an SDTrack test.
The method calibrates quantization to preserve the firing decisions of a floating-point reference model. Its objective accounts for how quantization errors affect firing thresholds and propagate across time steps. The authors test DAQ on image classification, semantic segmentation and event-based tracking. On the FE108 tracking benchmark with SDTrack, they report 58.8% AUC and 90.9% precision rate.
ArXiv · AI/CL/LG's note
The method calibrates quantization to preserve the firing decisions of a floating-point reference model. Its objective accounts for how quantization errors affect firing thresholds and propagate across time steps. The authors test DAQ on image classification, semantic segmentation and event-based tracking. On the FE108 tracking benchmark with SDTrack, they report 58.8% AUC and 90.9% precision rate.
ArXiv · AI/CL/LG's note
score 4