A robust approach to the design of machine learning algorithms, based on minimising finite sums of the parametrised functions is considered. This method implies using robust finite-sum differentiable aggregating functions that are resistant to outliers.