Methods similar to this, perhaps a little more elaborated had been in use in machine learning for around 20 years… sparse approximations (inducing points)