Learning on a Grassmann Manifold
Published:
In our latest paper, we have shown that the classical problem of beamforming codebook design in limited feedback MIMO systems is equivalent to finding K centroids of the projections of the channel realizations on a Grassmann manifold, which can be solved using an unsupervised learning technique, such as K-means clustering.
Since this problem is inspired by the large dimensionality of the channel matrices, the natural tendency is to think in terms of obtaining a lower dimensional representation of the channel using deep learning techniques. However, in this specific problem, we show that the analytical structure of the codebook design problem can be leveraged to arrive directly at the objective function of K-means clustering on the Grassmann manifold. The paper will be presented in Asilomar, 2020.