We consider data structured as attributed simplicial complexes, which are generalizations of graphs that include higher-dimensional simplices beyond vertices and edges; this structure allows for greater flexibility in modeling higher-order relationships. With this generalized framework, however, comes more complicated notions of message passing or convolution on the simplices to define simplicial neural network layers. While it is desirable in certain applications to use a more general input space of simplicial complexes, as opposed to graphs, such additional structure imposes high computational cost. To mitigate the increased computational cost as well as to reduce overfitting in the models, we propose a learned coarsening of simplicial complexes to be interleaved between standard simplicial neural network layers. This simplicial pooling layer generates hierarchical representations of simplicial complexes and is a generalized graph coarsening method which collapses information in a learned fashion. In this framework, a soft partition of vertices is learned at each layer and used to coarsen the simplicial complex for input to the subsequent layer. The pooling method extends learned vertex clustering to coarsen the higher dimensional simplices in a deterministic fashion. While in practice, the pooling operations are computed via a series of matrix operations, its topological motivation is based on unions of stars of simplices.