Latency-Constrained Input-Aware Quantization of Time Series Inference Workflows at the Edge

Quantization is a common technique for compressing Deep Neural Networks (DNNs) for edge device deployment, crucial for real-time time-series sensor data classification. While Adaptive Quantization (AQ) techniques allow updating the model's quantization scheme based on resource constraints without re-training, they are input-agnostic. Conversely, Input-aware Quantization (IQ) techniques adjust quantization per input, improving accuracy but ignoring resource constraints. Additionally, most quantization methods are designed for vision-based applications and underutilize time-frequency domain semantics. To this end, we present ReactQuant, an end-to-end framework for input-aware quantization of DNNs with strict resource constraints, specialized for time-series applications. ReactQuant features Partition-Derived Quantization, which segments time-series input based on its time and frequency domain features; a two-staged training process that restricts quantization space to each layer's promising bit-width candidates; and a novel algorithm to perform input-aware quantization within user-defined resource constraints. Extensive experimentation shows that ReactQuant outperforms AQ and IQ techniques in accuracy and ensures strict compliance with user-defined resource constraints. Under similar resource constraints, ReactQuant achieves up to 19% higher accuracy than AQ techniques and 4.6% higher than state-of-the-art IQ technique. It incurs no resource constraint violations, compared to up to 27.63% for state-of-the-art IQ technique.