This chapter introduces the background knowledge of structured representation learning which aims to enforce beneficial inductive biases to learned representations for improved generalization and robustness. Among the various valuable structures, we mainly focus on the literature of disentangled representations and equivariant networks. The historic developments and commonly used techniques of these two fields are first introduced, followed by an analysis of their existing limitations, resulting in restricted the real-world applications. At the intersection of these two fields, this book proposes a new promising research line – developing approximately equivariant neural networks through learning disentangled homomorphic representations.