Demystifying Inductive Biases for $\beta$-VAE Based Architectures