Zoom link: https://harvard.zoom.us/j/779283357?pwd=MitXVm1pYUlJVzZqT3lwV2pCT1ZUQT09
Passcode: 657361
Speaker: Michael Freedman
Title: What can ML learn from the proof of the Kolmogorov-Arnold theorem?
Abstract: The Kolmogorov-Arnold representation theorem shows that even very shallow, non-linear neural nets can express general continuous multivariate functions. I will begin by giving a proof. The theorem has often been regarded as "irrelevant" to machine learning because of the unrealistic precision required in its representation of Real numbers. I agree with this criticism but will present another path to ML-relevancy - not of the statement but of the proof.