Investigating generalization capabilities of neural networks by means of loss landscapes and Hessian analysis