University of Cambridge > Talks.cam > Artificial Intelligence Research Group Talks (Computer Laboratory) > Non-convex Optimisation Using the Polyak-Łojasiewicz Inequality

  Browse A-Z

Log in

University Account External (via Google)

🔐 Log In

Information on

Subscribing to talksdetails Finding a talkdetails Adding a talkdetails Disseminating talksdetails Help and Documentationdetails

ℹ️ Information

Non-convex Optimisation Using the Polyak-Łojasiewicz Inequality

Download to your calendar using vCal

If you have a question about this talk, please contact Mateja Jamnik .

We consider the idea of solving a non-convex optimisation problem by adding a large enough strongly-convex function to make the objective function convex. This allows the use of simpler convex optimisers, yet, given the large strong-convexity constant required, yields a value closer to the minimum of the function added than that of the objective one. We try to fix this with the novel idea of instead adding a function that makes the objective satisfy the Polyak-Łojasiewicz (PL) inequality, a much weaker condition than strong-convexity. Building on previous work, we construct an optimisation algorithm relying on this method. We find that a much smaller multiplicative constant is needed for convergence to a minimum. We attempt to find and prove convergence rates and computational complexity and test which algorithm yields a more accurate minimum.

This talk is part of the Artificial Intelligence Research Group Talks (Computer Laboratory) series.

This talk is included in these lists:

Note that ex-directory lists are not shown.

Other lists

You tube link to be confirmed Room A (East 1)/West Hub SG1 & SG2, Alison Richard Building, 7 West Rd, Cambridge CB3 9DT

Other talks

From Physics-informed Neural Networks to Numerically Verified Proofs in Geometric Analysis Title to be confirmed A flexible Monte Carlo quantitative bias analysis for unmeasured confounding TBC Title abstract tbc Afternoon Break
 

© 2006-2025 Talks.cam, University of Cambridge. Contact Us | Help and Documentation | Privacy and Publicity