Skip to content

Comment on Every Model Learned by Gradient Descent Is Approximately a Kernel Machine

Comments

From a skim, I think one asterisk which may be a gap between the claim in the title and what's actually shown in the paper is that the theorem focuses on gradient descent-trained models which minimize a loss function which is the sum of loss L(y_i, y*_i) on points from a given dataset. While that's clearly very broad, I think it _doesn't_ include things like GANs, where parts of the model produce fake data to train against.

The claim also applies to GANs as you can simply use a masking function to indicate which inputs were used for each optimization timestep, like the author suggests for stochastic gradient descent in remark 5.

AboutSource Built by g1lg1l

Hackerly is an independent reader for Hacker News, built on the public HN API. Not affiliated with Y Combinator.