TIDES: Implicit Time-Awareness in Selective State Space Models
TIDES keeps real time gaps intact while moving selectivity elsewhere in the state-space model.
The paper argues that Mamba-style selective SSMs gain expressivity by learning the time step, but that breaks alignment with irregular physical timestamps. TIDES instead leaves the step size equal to the observed time gap and puts input dependence on the diagonal state matrix. The authors say this lets it handle irregular time series natively while preserving per-token expressivity. They report stronger average results on UEA classification and Physiome ODE regression, plus competitive performance on six of eight irregular datasets. HF Daily Papers' note
The paper argues that Mamba-style selective SSMs gain expressivity by learning the time step, but that breaks alignment with irregular physical timestamps. TIDES instead leaves the step size equal to the observed time gap and puts input dependence on the diagonal state matrix. The authors say this lets it handle irregular time series natively while preserving per-token expressivity. They report stronger average results on UEA classification and Physiome ODE regression, plus competitive performance on six of eight irregular datasets. HF Daily Papers' note
score 5