We propose that learning in deep neural networks proceeds in two phases: a rapid curve fitting phase followed by a slower compression or coarse graining phase. This view is supported by the shared temporal structure of three phenomena: grokking, double descent and the information bottleneck, all of which exhibit a delayed onset of generalization well after training error reaches zero. We empirically show that the associated timescales align in two rather different settings. Mutual information between hidden layers and input data emerges as a natural progress measure, complementing circuit-based metrics such as local complexity and the linear mapping number. We argue that the second phase is not actively optimized by standard training algorithms and may be unnecessarily prolonged. Drawing on an analogy with the renormalization group, we suggest that this compression phase reflects a principled form of forgetting, critical for generalization.

We study the well-known grokking phenomena in neural networks (NNs) using a 3-layer MLP trained on 1 k-sample subset of MNIST, with and without weight decay, and discover a novel third phase -- \emph{anti-grokking} -- that occurs very late in training and resembles but is distinct from the familiar \emph{pre-grokking} phases: test accuracy collapses while training accuracy stays perfect. This late-stage collapse is distinct, from the known pre-grokking and grokking phases, and is not detected by other proposed grokking progress measures. Leveraging Heavy-Tailed Self-Regularization HTSR through the open-source WeightWatcher tool, we show that the HTSR layer quality metric $α$ alone delineates all three phases, whereas the best competing metrics detect only the first two. The \emph{anti-grokking} is revealed by training for $10^7$ and is invariably heralded by $α< 2$ and the appearance of \emph{Correlation Traps} -- outlier singular values in the randomized layer weight matrices that make the layer weight matrix atypical and signal overfitting of the training set. Such traps are verified by visual inspection of the layer-wise empirical spectral densities, and by using Kolmogorov--Smirnov tests on randomized spectra. Comparative metrics, including activation sparsity, absolute weight entropy, circuit complexity, and $l^2$ weight norms track pre-grokking and grokking but fail to distinguish grokking from anti-grokking. This discovery provides a way to measure overfitting and generalization collapse without direct access to the test data. These results strengthen the claim that the \emph{HTSR} $α$ provides universal layer-convergence target at $α\approx 2$ and underscore the value of using the HTSR alpha $(α)$ metric as a measure of generalization.

Grokking, referring to the abrupt improvement in test accuracy after extended overfitting, offers valuable insights into the mechanisms of model generalization. Existing researches based on progress measures imply that grokking relies on understanding the optimization dynamics when the loss function is dominated solely by the weight decay term. However, we find that this optimization merely leads to token uniformity, which is not a sufficient condition for grokking. In this work, we investigate the grokking mechanism underlying the Transformer in the task of prime number operations. Based on theoretical analysis and experimental validation, we present the following insights: (i) The weight decay term encourages uniformity across all tokens in the embedding space when it is minimized. (ii) The occurrence of grokking is jointly determined by the uniformity of the embedding space and the distribution of the training dataset. Building on these insights, we provide a unified perspective for understanding various previously proposed progress measures and introduce a novel, concise, and effective progress measure that could trace the changes in test loss more accurately. Finally, to demonstrate the versatility of our theoretical framework, we design a dedicated dataset to validate our theory on ResNet-18, successfully showcasing the occurrence of grokking. The code is released at https://github.com/Qihuai27/Grokking-Insight.

Neural networks trained to solve modular arithmetic tasks exhibit grokking, a phenomenon where the test accuracy starts improving long after the model achieves 100% training accuracy in the training process. It is often taken as an example of "emergence", where model ability manifests sharply through a phase transition. In this work, we show that the phenomenon of grokking is not specific to neural networks nor to gradient descent-based optimization. Specifically, we show that this phenomenon occurs when learning modular arithmetic with Recursive Feature Machines (RFM), an iterative algorithm that uses the Average Gradient Outer Product (AGOP) to enable task-specific feature learning with general machine learning models. When used in conjunction with kernel machines, iterating RFM results in a fast transition from random, near zero, test accuracy to perfect test accuracy. This transition cannot be predicted from the training loss, which is identically zero, nor from the test loss, which remains constant in initial iterations. Instead, as we show, the transition is completely determined by feature learning: RFM gradually learns block-circulant features to solve modular arithmetic. Paralleling the results for RFM, we show that neural networks that solve modular arithmetic also learn block-circulant features. Furthermore, we present theoretical evidence that RFM uses such block-circulant features to implement the Fourier Multiplication Algorithm, which prior work posited as the generalizing solution neural networks learn on these tasks. Our results demonstrate that emergence can result purely from learning task-relevant features and is not specific to neural architectures nor gradient descent-based optimization methods. Furthermore, our work provides more evidence for AGOP as a key mechanism for feature learning in neural networks.

Grokking, a phenomenon where machine learning models generalize long after overfitting, has been primarily observed and studied in algorithmic tasks. This paper explores grokking in real-world datasets using deep neural networks for classification under the cross-entropy loss. We challenge the prevalent hypothesis that the $L_2$ norm of weights is the primary cause of grokking by demonstrating that grokking can occur outside the expected range of weight norms. To better understand grokking, we introduce three new progress measures: activation sparsity, absolute weight entropy, and approximate local circuit complexity. These measures are conceptually related to generalization and demonstrate a stronger correlation with grokking in real-world datasets compared to weight norms. Our findings suggest that while weight norms might usually correlate with grokking and our progress measures, they are not causative, and our proposed measures provide a better understanding of the dynamics of grokking.

In-context learning is a powerful emergent ability in transformer models. Prior work in mechanistic interpretability has identified a circuit element that may be critical for in-context learning -- the induction head (IH), which performs a match-and-copy operation. During training of large transformers on natural language data, IHs emerge around the same time as a notable phase change in the loss. Despite the robust evidence for IHs and this interesting coincidence with the phase change, relatively little is known about the diversity and emergence dynamics of IHs. Why is there more than one IH, and how are they dependent on each other? Why do IHs appear all of a sudden, and what are the subcircuits that enable them to emerge? We answer these questions by studying IH emergence dynamics in a controlled setting by training on synthetic data. In doing so, we develop and share a novel optogenetics-inspired causal framework for modifying activations throughout training. Using this framework, we delineate the diverse and additive nature of IHs. By clamping subsets of activations throughout training, we then identify three underlying subcircuits that interact to drive IH formation, yielding the phase change. Furthermore, these subcircuits shed light on data-dependent properties of formation, such as phase change timing, already showing the promise of this more in-depth understanding of subcircuits that need to "go right" for an induction head.

Transformer models exhibit in-context learning: the ability to accurately predict the response to a novel query based on illustrative examples in the input sequence. In-context learning contrasts with traditional in-weights learning of query-output relationships. What aspects of the training data distribution and architecture favor in-context vs in-weights learning? Recent work has shown that specific distributional properties inherent in language, such as burstiness, large dictionaries and skewed rank-frequency distributions, control the trade-off or simultaneous appearance of these two forms of learning. We first show that these results are recapitulated in a minimal attention-only network trained on a simplified dataset. In-context learning (ICL) is driven by the abrupt emergence of an induction head, which subsequently competes with in-weights learning. By identifying progress measures that precede in-context learning and targeted experiments, we construct a two-parameter model of an induction head which emulates the full data distributional dependencies displayed by the attention-based network. A phenomenological model of induction head formation traces its abrupt emergence to the sequential learning of three nested logits enabled by an intrinsic curriculum. We propose that the sharp transitions in attention-based networks arise due to a specific chain of multi-layer operations necessary to achieve ICL, which is implemented by nested nonlinearities sequentially learned during training.

Neural networks often exhibit emergent behavior, where qualitatively new capabilities arise from scaling up the amount of parameters, training data, or training steps. One approach to understanding emergence is to find continuous \textit{progress measures} that underlie the seemingly discontinuous qualitative changes. We argue that progress measures can be found via mechanistic interpretability: reverse-engineering learned behaviors into their individual components. As a case study, we investigate the recently-discovered phenomenon of ``grokking'' exhibited by small transformers trained on modular addition tasks. We fully reverse engineer the algorithm learned by these networks, which uses discrete Fourier transforms and trigonometric identities to convert addition to rotation about a circle. We confirm the algorithm by analyzing the activations and weights and by performing ablations in Fourier space. Based on this understanding, we define progress measures that allow us to study the dynamics of training and split training into three continuous phases: memorization, circuit formation, and cleanup. Our results show that grokking, rather than being a sudden shift, arises from the gradual amplification of structured mechanisms encoded in the weights, followed by the later removal of memorizing components.