Test loss

See also: Machine learning terms

Introduction

Test loss is a metric that measures a model's loss against the test data set. Note that the test dataset is a separate dataset from the training data set and the validation data set. Testing the model on the test set is like a final test for an already trained machine learning model.

Machine learning algorithms measure their model's ability to make accurate predictions on unseen data. The test loss provides an assessment of a model's generalization ability, or its capacity for making accurate predictions when presented with new information that was not seen during training.

The test loss is calculated by comparing the model's predictions on test data with actual values for target variables (labels). This difference, known as an error, serves to measure how accurately predictions made on the test data reflect actual outcomes. It serves to reflect how well-fitted the model's predictions were to the actual data.

We want to minimize the test loss. A large test loss vs. training loss or validation loss might indicate that we are overfitting the model and might need to use regularization.

Calculating a test loss requires consideration of the particular problem being addressed and desired properties of the model. Common loss functions include mean squared error, mean absolute error, and categorical cross-entropy.

Mean Squared Error

Mean squared error (MSE) is a commonly used measure when attempting to predict an ongoing target variable. MSE is calculated as the average of squares between predicted values and actual values (labels).

MSE is a smooth and differentiable function, making it suitable for optimization algorithms such as gradient descent. Furthermore, MSE has the advantageous property of being sensitive to large errors; this means a model with an increased MSE is likely to make major mistakes in some instances in its test set.

Mean Absolute Error

Mean absolute error (MAE) is a commonly used measure for regression problems. MAE is calculated as the average of all residual values between predicted values and actual values.

MAE is a robust loss function that is insensitive to outliers, making it ideal for problems where there may be some instances in the test set with large errors. Unlike MSE, however, MAE is nondifferentiable which may make optimizing with gradient-based algorithms more challenging.

Categorical Cross-Entropy

Categorical cross-entropy is a widely used approach in classification problems, where the aim is to accurately predict a categorical target variable. Categorical cross-entropy is calculated as the average of negative log likelihoods associated with predicted class probabilities.

Categorical cross-entropy is a smooth and differentiable function with the desirable property of assigning a large loss to predictions with low confidence. This property makes categorical cross-entropy ideal for classification problems where it's necessary to penalize models for making incorrect predictions with high assurance.

Explain Like I'm 5 (ELI5)

The test loss is an indicator of how well a machine learning model can predict unknown events. It compares what the model thinks will occur with what actually does, and there are various methods for calculation depending on the problem at hand. If there are many mistakes made by the model, its test loss will be high.