Customized Mamba SSM Block Deep Learning Layer

Core Mamba SSM block implementation leveraging selective state spaces based on MATLAB Deep Learning Toolbox

This implementation presents the core Mamba SSM Block, the fundamental building unit of Mamba sequence models. Different from conventional SSMs, its state transition parameters are dynamically conditioned on input features, supporting adaptive sequence modeling. This lightweight block features linear computational complexity, well-suited for long sequence processing in time-series, NLP and related applications.

It should be noted that this layer is a subcomponent of the integrated mamba block. To implement a complete mamba module, the users can combine this layer with other builtin deep learning layers, e.g., convolution1dLayer, fullyConnectedLayer, multiplicationLayer, sigmoidLayer, and connect these layers according to the topology structure shown in the figure. For more details about the mamba block, the users can refer to Mamba: Linear-Time Sequence Modeling with Selective State Spaces