Enhancing deep learning algorithm accuracy and stability using multicriteria optimization: an application to distributed learning with MNIST digits

The training phase is the most crucial stage during machine learning process. In case of labeled data and supervised learning, entails minimizing loss function under various constraints. We provide an innovative model for with numerous sets, resulting from application multicriteria optimization techniques to existing deep algorithms. Data fitting formulated as a in which each criterion measures error on specific set. This involving vector-valued function, it has be analyzed using notion Pareto efficiency. present stability results efficient solutions presence input output perturbations. multiple set environment comes into play eliminate bias caused by selection To apply this concept, we scalarization strategy well numerical experiments digit classification MNIST data.