This paper introduces Temporal Adaptive Bitwise Soft Quantization (TABSQ), a novel method for compressing temporal sensor data streams. TABSQ extends differentiable soft quantization to time-series data with dynamically allocated bit-widths, addressing challenges in temporal dynamics, streaming constraints, and heterogeneous precision requirements.
Key findings
TABSQ combines a temporal-aware soft quantization module, adaptive bit allocation controller, and a reconstruction-aware training objective.
The method uses Gumbel-Softmax relaxation for end-to-end learning of quantization parameters.
A temporal attention-based bit allocation controller dynamically assigns bit-widths based on local temporal context and predicted information content.
TABSQ achieves superior compression-efficiency trade-offs compared to static quantization baselines.
Limitations & open questions
The paper is a research proposal and does not include experimental results.
The effectiveness of TABSQ in real-world IoT deployments remains to be validated.