转载：Why machine learning algorithms are hard to tune and how to fix it

Key point

• We often optimize a Linear combinations of losses and hope to simultaneously reduce both $L_1$ and $L_0$ losses, but that this linear combination is actually precarious and treacherous.
• Authours showed that when the pareto curve is concave, only one of these two losses was considered and this linear combination was valid in convex pareto curve.
• In fact, We can't figure out the property of the pareto curve.
• We can reformulate the linear combination losses as a constraint optimization problem, e.g., restricting $L_0\le ϵ$
• By doing so, authors gave three possible solutions.

References

• https://www.engraved.blog/why-machine-learning-algorithms-are-hard-to-tune
• https://www.engraved.blog/how-we-can-make-machine-learning-algorithms-tunable