Modern systems neuroscience benefits from the ability to record from and digitize a large amount of functional data from hundreds or even thousands of neurons. Understanding, transmitting, storing, and parsing information of such volume and complexity calls for methods of dimensionality reduction. One observation about neuronal activity in mammalian brains is that populations are sparsely active; that is, only a small subset of the whole ensemble is coactive at any moment. This property may be exploited to summarize information content succinctly. This paper tests the hypothesis that information contained in ensemble activity recorded from the primate motor cortex about limb movements is preserved when the activity is projected onto a sparse basis. Spiking rate data from neurons in the motor cortex of an awake behaving macaque monkey was compressed using a sparse autoencoder network, and classifications of movement directions were made in the compressed space. Classifier performance is shown to be similar when using either compressed (sparsened) or uncompressed neural activity, demonstrating the potential use of the sparse autoencoder as an unsupervised compression algorithm for low power/low bandwidth wireless transmission of neural ensemble data.