Gradient Agreement as an Optimization Objective for Meta-Learning