STAMP: Shared-Dictionary Shapelet Tokenization For Multi-Resolution Time-Series Compression

Wearables, mobile devices, and Internet of Things platforms stream multimodal time series continuously and at scale, making storage and transmission a pri- mary bottleneck under tight bandwidth, memory, and energy budgets. Meanwhile, downstream pipelines, including token based large-model workflows, benefit from discrete, rate controlled representations that are easy to serve and process. A core challenge is that time-series semantics are inherently multi-resolution, spanning sub-second transients and minute to hour scale structure, and are further degraded by small temporal misalignment from jitter and imperfect segmentation. Existing discretization and tokenization methods often rely on fixed stride segmentation or generic primitives, which can either miss short events or become inefficient over long horizons. To address this, we present STAMP, a multi-resolution neu- ral compression framework built around a shared dictionary of learned tempo- ral primitives. STAMP encodes signals using a coarse to fine residual cascade, discretizing residuals via vector quantization to produce a compact discrete code sequence with controllable rate. A shift robust additive decoder reconstructs the signal by superposing translated dictionary atoms across scales, improving robust- ness to boundary shifts while keeping decoding lightweight. We evaluate STAMP on a synthetic multimodal benchmark (SeqComb), measuring rate, distortion, and perceptual fidelity trade-offs alongside end-to-end inference latency and peak syn- thesis memory, and comparing against fundamental baselines. In this controlled setting, STAMP achieves favorable rate distortion behavior while preserving local morphology at low decoding cost, suggesting a practical path toward edge deploy- able discrete representations for large scale time-series streams.