Semi-supervised transfer learning with genetic algorithm tuned transformation and novel label transfer mechanism

Robotics and intelligent sensing methods are experiencing a new wave applications through the use of machine learning systems. Intelligence is being introduced in robots and sensor platforms by utilizing machine learning techniques such as classification. In the field of robotics, generating training data can be very complex and often, expensive. In this set-up, transfer learning can greatly improve the performance of a classifier wherever and whenever enough labeled data is not available in a domain of interest (target domain), but ample labeled data can be found in a different but related domain (source domain). A new optimized method is proposed in this work to transform the observation from source domain along with a new label transfer mechanism. The transformed, or adapted, domain has the same number of features as the target domain and the same number of observations from the source domain. Labels are transferred from source to target domain using a multivariate Gaussian mixture model (GMM). Genetic algorithm is used to optimize the transformation process by minimizing a cost function that addresses both distribution difference and accuracy. Experiments show that the proposed method outperforms any classifier trained only with source or target domain data.