The empirical success of deep learning presents numerous challenges to theoreticians. In literature, four main factors are addressed:
Architectures (5 papers)
Optimization Algorithms (3 papers)
Generalization/Regularization Techniques (3 papers)
Stability/Robustness (6 papers)
Understanding the necessity and interplay of these three factors is essential in an analysis of their success. Thus, the theoretical aspects mainly concern:
Interpretability/ Understanding (3 papers)
Provability/Stability (3 papers)
Adversarial Perturbations (19 papers)
Noisy Labels (2 papers)