STLn

STLn stands for Spatiotemporal Linear Network, a class of models designed to capture both spatial structure and temporal dynamics in sequential data. Originating in the context of machine learning for time-series and video data, STLn refers to networks that emphasize linear transformations over short temporal windows, while allowing stochasticity through noise terms or probabilistic latent variables.

Conceptually, an STLn layer applies a linear spatial transform followed by a linear temporal update, often implemented as a state-space or linear recurrent component. The core idea is to maintain interpretability and computational efficiency by avoiding heavy nonlinearities within the core dynamics, while optionally integrating nonlinear readouts or attention mechanisms.

Training: STLn models are trained via backpropagation through time or truncated BPTT, using objectives such as mean-squared error for forecasting or reconstruction. Regularization through noise injection, variational terms, or sparsity penalties is common. The linear structure allows efficient inference and stable gradients on long sequences.

Applications and evaluation: STLn models have been explored for video frame prediction, traffic and climate forecasting, and sensor networks. They are praised for tractability and speed, but may underfit highly nonlinear phenomena unless augmented with selective nonlinear components.

Relation to other work: STLn relates to linear dynamical systems, state-space models, and recurrent neural networks. It sits at the intersection of classical time-series modeling and modern deep learning, offering a middle ground between expressiveness and interpretability.