Quantizing Time-Series Models As Dynamical Systems: Trajectory-Based Quantization Sensitivity Score
Researchers have introduced the Trajectory-based Quantization Sensitivity Score (TQS), a novel metric for evaluating post-training quantization (PTQ) in time-series models. TQS reframes quantization as a dynamical systems problem, assessing how errors propagate and amplify over time. This approach allows for a priori sensitivity estimation, independent of specific quantizer selection or bit-width, enabling better budget planning even for complex, compiled networks. The TQS-PTQ framework, which requires no calibration data or complex approximations, demonstrates robust performance for low-precision deployment in resource-constrained environments. AI
IMPACT This new quantization sensitivity metric could enable more efficient deployment of AI models in resource-constrained environments.