As part of enhancing Linear Learner, Amazon SageMaker has added automatic early stopping with or without a validation dataset during model tuning. If you provide the Linear Learner algorithm with a validation data set, your model training will stop early once validation loss stops improving. If no validation set is available, model training will stop early when training loss stops improving and reverts to the best model.
In addition, there are several new ways to customize the Linear Learner algorithm for your model training using new loss functions for the Linear Learner hyperparameters. You can now use eight new loss functions in Amazon SageMaker: Squared Loss, for most regression problems, to estimate the mean; Absolute Loss, to produce estimates of the median; Quantile Loss, to provide a quantile value on which to make predictions (e.g. 0.9 quantile of the distribution); Huber Loss, for training with Squared loss but avoiding outlier sensitivity; Epsilon-Insensitive Loss, for specifying a threshold for acceptable errors; Logistic Regression, for binary classification problems; Hinge Loss, also known as Support Vector Machine (SVM) for binary classifications. Finally, Linear Learner also allows you to specify class weights, for highly unbalanced training data in binary classification problems.
These new features in Amazon SageMaker are available today in the U.S. East (N. Virginia), U.S. East (Ohio), EU (Ireland), and U.S. West (Oregon) AWS Regions. For additional details for the Linear Learner enhancements, including the new loss functions, visit AWS Machine Learning Blog.
