Machine Learning Systems Design with Sara Hooker: Robustness

The transcript discusses the importance of focusing on data rather than just models for interpretability. It explores simplifying inputs to improve model performance and the trade-offs involved. Curriculum learning and model signals are examined to distinguish between easy and hard examples. The discussion highlights the significance of understanding relative importance in interpretability and the challenges of handling uncertainty, particularly distinguishing between noise and rare examples.

In what ways can model signal be used to improve training outcomes?

What role does understanding relative importance play in model interpretability?

How can we leverage early training signals to distinguish between noisy and rare examples?

What are the implications of distinguishing between noise and rare artifacts in machine learning?