Learning identifiable and interpretable latent models of high-dimensional neural activity using pi-VAE