Green Machine Learning via Augmented Gaussian Processes and Multi-Information Source Optimization

Searching for accurate Machine and Deep Learning models is a computationally\nexpensive and awfully energivorous process. A strategy which has been gaining\nrecently importance to drastically reduce computational time and energy\nconsumed is to exploit the availability of different information sources, with\ndifferent computational costs and different "fidelity", typically smaller\nportions of a large dataset. The multi-source optimization strategy fits into\nthe scheme of Gaussian Process based Bayesian Optimization. An Augmented\nGaussian Process method exploiting multiple information sources (namely,\nAGP-MISO) is proposed. The Augmented Gaussian Process is trained using only\n"reliable" information among available sources. A novel acquisition function is\ndefined according to the Augmented Gaussian Process. Computational results are\nreported related to the optimization of the hyperparameters of a Support Vector\nMachine (SVM) classifier using two sources: a large dataset - the most\nexpensive one - and a smaller portion of it. A comparison with a traditional\nBayesian Optimization approach to optimize the hyperparameters of the SVM\nclassifier on the large dataset only is reported.\n