1) Norm(weights, N) will return (w_1^N + w_2^N +....)^(1/N), so norm * norm is required.
2) This is bug as you said. I intend to fix this using weighted regularization, and intercept term will be regularized with weight zero. https://github.com/apache/spark/pull/1518 But I never actually have time to finish it. In the meantime, I'm fixing this without this framework in new ML pipeline framework. 3) I think in the long term, we need weighted regularizer instead of updater which couples regularization and adaptive step size update for GD which is not needed in other optimization package. Sincerely, DB Tsai ------------------------------------------------------- Blog: https://www.dbtsai.com On Tue, Apr 7, 2015 at 3:03 PM, Ulanov, Alexander <alexander.ula...@hp.com> wrote: > Hi, > > Could anyone elaborate on the regularization in Spark? I've found that L1 and > L2 are implemented with Updaters (L1Updater, SquaredL2Updater). > 1)Why the loss reported by L2 is (0.5 * regParam * norm * norm) where norm is > Norm(weights, 2.0)? It should be 0.5*regParam*norm (0.5 to disappear after > differentiation). It seems that it is mixed up with mean squared error. > 2)Why all weights are regularized? I think we should leave the bias weights > (aka free or intercept) untouched if we don't assume that the data is > centralized. > 3)Are there any short-term plans to move regularization from updater to a > more convenient place? > > Best regards, Alexander > --------------------------------------------------------------------- To unsubscribe, e-mail: dev-unsubscr...@spark.apache.org For additional commands, e-mail: dev-h...@spark.apache.org
