On the Fault Tolerance of Self-Supervised Training in Convolutional Neural Networks