This data dictionary enables the specification of metadata that describes products generated by the use of a machine learning model. It captures information about how the model was trained and evaluated. ## CHANGE LOG ## 1.2.0.0 - Add several new classes: Optimization Algorithm; Augmentation Method; Preprocessing Method; Trained Model Hyperparameter; Trained Model Architecture; Calibration Method; Trained Model File. - Add trained_model_creation_date attribute to Trained Machine Learning Model class. - Add trained_model_objective_function_name and trained_model_objective_function_description to Machine Learning Algorithm class. 1.1.0.0 - Change data_set_size to data_set_count. - Expand attribute definitions. - Update steward name. 1.0.1.0 - Address Oxygen-flagged errors. - Make data_set_size a ASCII_NonNegative_Integer. - Update some attribute and class definitions. - Update README file per latest template. - Use PDS Information Model 1.18 (I). 1.0.0.0 - Initial release. The Augmentation_Method class contains information about a data set's augmentation method. The Calibration_Method class is a container for classes and and attributes describing the Calibration Method. An external reference to a citation, such as a DOI, is optional. The Data_Set class is the container for classes and attributes describing the size and version of data sets used by the machine learning model. The Machine_Learning class is a container for all machine learning information in the label. The Machine_Learning_Algorithm class is a container for classes and and attributes describing the algorithm type and learning style used. An external reference to a citation for the algorithm is required. The Optimization_Algorithm class is a container for classes and and attributes describing the optimization algorithm used. The Preprocessing_Method class contains information about a data set's preprocessing method. The Test_Performance class contains information about a trained model's performance on the test set. The Test_Set class belongs to the Data_Set class family and contains attributes describing the size and version of the data set used to test the machine learning model (i.e., in terms of generalization to previously unseen data). The Trained_Machine_Learning_Model class is a container for information about how a given model was trained and evaluated. Machine_Learning_Algorithm, Training_Set, Optimization_Algorithm, Trained_Model_Hyperparameter, and Trained_Model_File are required, while Validation_Set, Test_Set, Test_Performance, Trained_Model_Architecture, and Calibration_Method are optional. The Trained_Model_Architecture class is a container for classes and and attributes describing the trained model architecture. An internal reference to the architecture is optional. The Trained_Model_File class is a container for classes and and attributes describing the trained model file. An external reference for the trained model file is required. The Trained_Model_Hyperparameter class is a container for classes and and attributes describing the trained model hyperparameter. The Training_Set class belongs to the Data_Set class family and contains attributes that describe the size and version of the data set used to train the machine learning model. The Validation_Set class belongs to the Data_Set class family and contains attributes that describe the size and version of the data set used to validate the machine learning model (e.g., to choose the best hyperparameters). This section contains the simpleTypes that provide more constraints than those at the base data type level. The simpleTypes defined here build on the base data types. This is another component of the common dictionary and therefore falls within the common namespace. The algorithm_learning_style attribute describes the type of learning style employed by the algorithm to solve a problem. Specifically, the learning style depends on whether labeled or unlabeled data was employed to train the model. Labeled data includes observations that are associated with a desired output such as a class or numeric value. The algorithm_name attribute specifies the name of the algorithm used. The algorithm_type attribute describes the kind of algorithm used, such as a regression model, neural network, tree, etc. The augmentation_method_description attribute describes the augmentation method in more detail. The augmentation_method_name attribute names any method by which an image was augmented, e.g. 180 degrees counterclockwise rotation; horizontal flipping; brightness adjustment. The augmentation_method_parameter attribute names any parameters, e.g. gamma. The augmentation_method_parameter_value attribute specifies the value of the named parameter, e.g. for a parameter "gamma," the value could be [0.5, 1.5]. The calibration_method_description attribute specifies any additional details about the calibration method used. The calibration_method_name attribute specifies the name of the calibration used, e.g. "temperature scaling." The calibration_method_parameter attribute specifies the name of the parameter used, e.g. "temperature." The calibration_method_parameter_value attribute sepcifies the value of the named parameter, e.g. for a parameter "temperature," the value could be "1.31." The data_set_count attribute provides the number of items in the data set. The data_set_version_id attribute specifies the data set version number. The optimization_algorithm_description attribute specifies any additional details about the optimization algorithm. The optimization_algorithm_name attribute specifies the name of the operation algorithm, e.g. "stochastic gradient descent." The optimization_algorithm_parameter attribute specifies the name of a parameter used, e.g. "conv1 learning rate," or "weight decay." The optimization_algorithm_parameter_value attribute specifies the value of the named parameter, e.g. for a parameter "conv1 learning rate," the value could be "0.001." The performance_measure attribute specifies the name of the measure (or metric) used to report performance of the model on the test set. The performance_score attribute reports the numeric score the model achieved using performance_measure on the test set. Values are not constrained since the measure may not be a strict metric. Examples could include accuracy, loss, runtime, memory consumption, etc. The preprocessing_method_description attribute describes the preprocessing method in more detail. The preprocessing_method_name attribute names any method by which an image was preprocessed, e.g. center cropping. The preprocessing_method_parameter attribute names any preprocessing parameters. The preprocessing_method_parameter_value attribute specifies the value of the named preprocessing parameter. The trained_model_architecture_description attribute specifies any additional details about the architecture. The trained_model_architecture_name attribute specifies the name of the architecture, e.g. "MSLNet." The trained_model_architecture_version_id attribute specifies the architecture version number. The trained_model_creation_date attribute specifies the date that the model was created, in Year/Month/Day format. The trained_model_file_description attribute specifies any additional details about the trained model file, e.g. the file format. The trained_model_file_name attribute specifies the name of the trained model file. The trained_model_hyperparameter_name attribute specifies the name of the hyperparameter, e.g. "batch size," "normalization method," "number of epochs," "conv1 dropout rate," "conv2 dropout rate." The trained_model_hyperparameter_value attribute specifies the value of the hyperparameter named, e.g."z-score" for normalization method, "100" for "number of epochs." The trained_model_name attribute specifies the name of the model used. The trained_model_objective_function_description attribute describes the objective function used. The trained_model_objective_function_name attribute specifies the name of the objective function used. The trained_model_version_id attribute specifies the trained model version number.