Leveraging sparse and shared feature activations for disentangled representation learning