I’m working on training an AI model and want to improve its interpretability. I’ve read about the concept of monosemanticity, which suggests that each unit (like a neuron or embedding) should correspond to a single, clear concept. How does enforcing this property improve model interpretability, and what are the common methods or algorithms used to encourage monosemantic representations during training?