Generalization capabilities of translationally equivariant neural networks