Fields
Pick a domain and keep going deeper: field, subfield, topic. Every step lists the papers at that point or below it. The counts are papers that are accepted or in review; papers that were not accepted are in the full library.
Subfield
Generalisation, double descent and grokking
Benchmarks: Deep double descent protocol, Modular arithmetic grokking, PGDL, Generalisation-measure study and others. Datasets: CIFAR-10 / CIFAR-100, MNIST, SVHN, CIFAR-10N / CIFAR-100N. Baselines: Deep double descent, Interpolation and the classical bias-variance trade-off, Neural Tangent Kernel, Non-vacuous PAC-Bayes bounds and others.