Device for transfer learning between modified tasks

摘要

Device for transfer learning of hyperparameters of a machine learning algorithm, the device comprising a machine-readable storage medium on which instructions are stored which, when executed by a computer, caused the computer to execute a method comprising the following steps: i) Provision of a current search space (X _i ) and a previous search space (X _i-1 ), a cost function (f) to be optimized, and a ranking (C _i-1 ) of evaluated hyperparametric configurations of a previous optimization step using the previous search space (X _i-1 ) with regard to the cost function (f), wherein the search spaces (X _i, X _i-1 ) are each defined on the basis of predetermined value ranges of the hyperparameters; i) Creating a reduced search area ( ) where the value ranges of the hyperparameters of the reduced search space ( ) corresponds to the value ranges of the hyperparameters of the current search space (X _i ), restrictedly dependent on the values of a predeterminable number (a) of the best hyperparameter configurations from the ranking (C _i-1 ); iii) Multiple, random pulling of candidate configurations from the reduced search space ( ) and the current search space X _i and applying the machine learning algorithm, each parameterized with the candidate configurations for optimizing the cost function (f); iv) creating a Tree Parzen Estimator (TPE) depending on the candidate solutions and the results of the machine learning algorithm applied to the cost function (f) to be optimized; v) repeating a drawing of further candidate configurations by means of the TPE from the current search space (X _i ) and applying the machine learning algorithm parameterized with the candidate configurations to the cost function; vi) creating a new ranking (C _i ) of the further candidate configurations and selecting a configuration from the new ranking (C _i ) as a hyperparametric configuration for the machine learning algorithm.