Unlocking State Space Models: Continuous, Recursive, Convolutional Insights

State Space Models (SSMs) offer continuous, recursive, and convolutional perspectives, enabling efficient training and inference through discretization methods like the HiPPO matrix, showcasing strong performance across domains.

• State Space Models (SSMs) have three views: continuous, recursive, and convolutional, each with its own advantages and tradeoffs.
• SSMs can be discretized from the continuous view, enabling efficient training via convolutions and efficient inference via recursions.
• Key aspects of SSMs include the discretization method and parameterization of the state matrix, with approaches like the HiPPO matrix showing strong performance on benchmarks across various domains.