Grokking: generalization long after memorization
Train a small network on an algorithmic task (modular arithmetic is the classic) and watch two curves. Training accuracy hits 100% quickly — the model has memorized. Test accuracy stays at chance for a long time... and then, thousands of steps later, with no change in the recipe, it suddenly snaps to perfect. This delayed phase transition is grokking, and it is the most vivid proof that fitting the training set and understanding the task are different events.
The emerging explanation ties grokking to the implicit bias of Guide 3 and the regularizer in the room: there is a memorizing solution (high norm, found fast) and a generalizing solution (low norm, structured — the network literally learns a Fourier-like circuit for modular addition). Weight decay slowly drains norm from the memorizing solution until the implicit and explicit regularization tip the balance toward the structured one. Grokking is double descent's cousin in time: the second descent, dramatically delayed.
Weight decay adds a norm penalty, so the low-norm generalizing solution eventually wins over the fast high-norm memorizing one — the engine behind grokking.
The information bottleneck view of learning
A different frame asks what a representation keeps and throws away. The information bottleneck theory models a layer's representation T as trading two quantities: how much it remembers about the input X versus how much it tells you about the label Y. A good representation is a minimal sufficient statistic — maximal information about Y, minimal about X. Training, on this view, has two phases: a fast fitting phase that grabs label information, then a slow compression phase that forgets irrelevant input detail.
The information bottleneck objective: compress the input (shrink I(X;T)) while keeping what predicts the label (grow I(T;Y)), traded off by β.
It is a beautiful, controversial story. Connecting it to generalization is intuitive — a representation that forgets input noise should transfer the held-out signal — and it links naturally to entropy and KL divergence as the native currency. But the empirical claims are contested: the compression phase depends heavily on how you estimate mutual information and on the activation function, and it does not always appear. Treat the information bottleneck as a generative hypothesis that sharpened the field's questions, not as settled fact.
Why depth, provably
Stepping from dynamics to expressivity: why deep rather than just wide? Depth separation results make this precise. There exist functions a deep network represents with a modest number of units that any shallow network needs exponentially many units to approximate — composition of nonlinearities folds space in a way width alone cannot cheaply replicate. The canonical examples build highly oscillatory functions whose number of linear regions grows exponentially in depth but only polynomially in width.
An input layer feeding two hidden layers and an output layer through weighted edges.
Read this against Guide 2's universal approximation: a single wide layer can approximate any function, but the theorem is silent on cost. Depth separation supplies the missing economics — depth buys exponential parameter efficiency for compositional, hierarchical targets, which is exactly the structure of natural data. It is the theoretical backbone behind 'deep' in deep learning, and it explains why the hierarchical representations networks build are not a luxury but the source of their efficiency.
Why the classical bounds still fail us
It is worth being blunt about the state of rigorous generalization theory, because it frames where the frontier is. The classical complexity measures — VC dimension, Rademacher complexity — scale with parameter count and so predict, for modern networks, generalization bounds larger than 1: completely vacuous. They cannot even rule out the memorizing solution. Refinements help but do not yet close the gap.
The classical VC bound: true risk ≤ empirical risk plus a complexity term in d_VC/n — vacuous for overparameterized nets where d_VC ≫ n.
- Margin-based bounds: normalize by the margin rather than count parameters — capacity is about how confidently, not how many.
- PAC-Bayes bounds: PAC-Bayes measures distance from a prior, and gives some of the only non-vacuous bounds for real networks — flat minima reappear here as low KL to a broad prior.
- Norm- and data-dependent measures: tie capacity to the trained weights and the actual data spectrum, echoing the benign-overfitting story of Guide 1.
Pull back and the shape of the field is clear. No single theory yet explains deep learning; instead a constellation of partial results — kernels, implicit bias, landscape geometry, information, depth, refined bounds — each illuminates one facet. The honest frontier is unification: turning these compatible-but-separate stories into one quantitative account of when, and how well, a given network will generalize. That open problem is the reason this is a research field and not a closed chapter.