Why it matters
Landscape shape determines algorithm choice. Understanding shapes ML + optimization intuition.
Advertisement
The architecture
Convex: any local min is global min.
Non-convex: many local minima; saddle points.
Advertisement
How it works end to end
Convex: LP, QP, some ML (logistic regression).
Non-convex: neural networks. Saddle points more common than local minima in high dim.
Techniques: momentum + noise escape saddles.