Sparse, Geometric Autoencoder Models of V1