Most deep learning models for time series treat data as a sequence of discrete tokens. Whether it is a Transformer or a Recurrent Neural Network (RNN), these architectures assume that the "tick" of the clock between observations is constant. If you train a model on data sampled at 100Hz and deploy it on a sensor running at 500Hz, the model typically fails; it perceives the world as moving five times slower than it actually is.
A growing body of research suggests a more robust alternative: modeling the underlying continuous-time signal rather than the discrete sequence. By parameterizing dynamics as differential equations or continuous kernels, these models achieve sampling-rate invariance, allowing them to generalize to unseen temporal resolutions without retraining.
The Problem with Discrete Time
Traditional models like RNNs learn a transition function $h_t = f(h_{t-1}, x_t)$. This function is "blind" to the actual time elapsed between $t-1$ and $t$. As noted by Sebastián Gómez-González and Hernán F. García, models defined over continuous input spaces exhibit superior robustness to sampling rate variations compared to these discrete-time counterparts. When the sampling frequency shifts, discrete models experience significant performance degradation because the temporal dependencies they learned during training no longer align with the input stream.
The Discretization Lever
The breakthrough in making these models practical for deep learning came with the Linear State-Space Layer (LSSL), introduced by Albert Gu and colleagues in Combining Recurrent, Convolutional, and Continuous-time Models with Linear State Space Layers.
The core idea is to define the system's evolution using a continuous differential equation: $$\dot{h}(t) = \mathbf{A}h(t) + \mathbf{B}x(t)$$ To process digital data, this continuous system is discretized using a step size $\Delta$. The crucial insight is that $\Delta$ is a tunable parameter. If the sampling rate changes at inference time, we can simply scale $\Delta$ to compensate.
This mechanism provides a "mathematical zoom lens." As Nikola Zubić and Mathias Gehrig demonstrate in State Space Models for Event Cameras, adjusting this discretization parameter allows models to maintain high accuracy even when an event camera's frequency fluctuates wildly during deployment. Similarly, Arnab Kumar Mondal shows in HumMUSS that this allows human motion models to generalize to unseen video frame rates zero-shot.
Continuous Kernels and Basis Functions
Beyond State Space Models (SSMs), researchers are exploring other ways to parameterize time continuously.
In PLEIADES, Yan Ru Pei and Olivier JMD Coenen construct temporal convolution kernels using orthogonal polynomials. Instead of learning a fixed vector of weights, the model learns the coefficients of a continuous function. To perform inference at any sampling rate, the model simply samples this function at the required intervals.
Another approach, FlowState by Lars Graf and Stanisław Woźniak, uses a "Functional Basis Decoder." By treating time series as continuous signals, FlowState achieves "sampling-rate equivariance," outperforming models 192× its size when forced to adapt to different time scales on the fly.
Zero-Shot Robustness in the Real World
The practical implications of this research are significant for domains where hardware is heterogeneous or data is messy:
- Medical Diagnostics: Nils Strodthoff and Temesgen Mehari found that SSMs trained on 100Hz ECG data can be deployed on high-frequency hardware without retraining, a property they call "zero-shot robustness" (Advancing the State-of-the-Art for ECG Analysis through Structured State Space Models).
- Environmental Sensing: In Early Detection of Extreme Storm Tide Events, Marcel R. de Barros and colleagues use continuous temporal representations to handle irregularly sampled sensor data and missing values without needing complex imputation.
- Stochastic Control: Cristopher Salvi and others demonstrate that Neural Rough Differential Equations (Neural RDEs) can be trained on sparse data and yet remain accurate when evaluated on much finer time scales (A Neural RDE approach for continuous-time non-Markovian stochastic control problems).
By shifting the focus from "sequences of numbers" to "underlying continuous dynamics," these models move closer to how physical systems actually operate, gaining the ability to ignore the arbitrary clock speed of the sensors observing them.