Pith. sign in

REVIEW 3 major objections 6 minor 48 references

An In-depth Investigation of Sparse Rate Reduction in Transformer-like Models

T0 review · 3 major / 6 minor · reviewed 2026-08-12 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Sparse Rate Reduction, the objective behind the CRATE Transformer-like model, predicts generalization better than 21 baseline complexity measures and works as a regularizer.

desk verdict Solid, honest CRATE investigation with a real artifact finding and a plausible but protocol-sensitive generalization correlation that needs a sensitivity check before being called predictive. read the letter →

arxiv 2411.17182 v1 pith:W7OIS7IE submitted 2024-11-26 cs.LG

Yunzhe Hu , Difan Zou , Dong Xu This is my paper · ORCID
classification cs.LG
keywords sparseratereductionCRATEgeneralizationgapcomplexitymeasurealgorithmunrollingKendallrankcorrelationcodingregularization
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper sets out to show that Sparse Rate Reduction (SRR) — the information-theoretic objective that the white-box Transformer-like CRATE architecture is designed to optimize — is more than a design story: it is a usable predictor of generalization and a usable regularizer. On a collection of 64 CRATE variants obtained by varying four hyperparameters and four implementations, the layer-averaged SRR measure ranks models by generalization gap with Kendall's $\tau = 0.445$ at width 384 (0.407 at width 768), ahead of all 21 norm-, margin-, and sharpness-based baselines tested. Adding a stop-gradient SRR penalty to the cross-entropy loss improves top-1 accuracy on CIFAR-10/100 by small, consistent margins. Along the way the paper identifies a flaw in the original CRATE derivation: the self-attention-style update with a skip connection that was supposed to compress representations actually increases the subspace coding rate, because the second-order Taylor expansion used to derive it drops the first-order term. The reason to care is that if SRR really tracks generalization, an architecturally interpretable objective gains a practical role in model selection and training.

What carries the argument

The load-bearing object is the SRR complexity measure in equation (11), the layer average of $\lambda \|Z^{\ell}\|_0 + R_c(Z^{\ell};U^{\ell}) - R(Z^{\ell})$, where $R(Z) = \frac{1}{2} \log \det\left(I + \frac{d}{N\epsilon^2}Z^{\top}Z\right)$ is the coding rate of the token matrix and $R_c(Z;U) = \sum_{k} R(U_k^{\top}Z)$ is the coding rate of tokens projected onto $K$ incoherent low-dimensional subspaces spanned by $U$. The second key object is the MSSA update with skip connection, equation (8), $Z \leftarrow Z + \alpha\gamma^2 \sum_k U_k U_k^{\top}Z\,\mathrm{softmax}((U_k^{\top}Z)^{\top}(U_k^{\top}Z))$, which the paper derives from a Taylor expansion of the gradient of $R_c$ and then shows acts as an ascent step on $R_c$. These two objects carry the argument: the paper measures SRR on the forward passes of four CRATE implementations, correlates it with the generalization gap using Kendall's $\tau$, and uses the same measure, with gradients stopped, as a regularizer in equation (14).

What would settle it

Compute the SRR measure on the same 64 trained models with the Layer Normalization modules left in the forward pass (e.g., folding their scale and shift parameters into the coding-rate computation), and recompute Kendall's $\tau$ against the generalization gap; if the correlation falls below the sharpness-based baselines, the reported $\tau = 0.445$ is an artifact of the LayerNorm-removal protocol. A complementary check is to train additional CRATE variants at hyperparameter settings outside the original 64 combinations (for example, learning rate $5\times 10^{-5}$ or width 576) and test whether the in-sample correlation rank order persists.

Watch

Extended reading notes

Core claim

The paper's central claim is that SRR, evaluated as a layer-averaged complexity measure on the forward pass of CRATE-family models, has genuine predictive power for generalization and can be used to improve training. Concretely, the authors train 64 models spanning CRATE-C, CRATE-N, CRATE-T, and CRATE with varied batch size, learning rate, width, and dropout; they report that the measure $\mu_{\mathrm{SRR}}(w;Z) = \frac{1}{L} \sum_{\ell} [\lambda \|Z^{\ell}\|_0 + R_c(Z^{\ell};U^{\ell}) - R(Z^{\ell})]$ has Kendall's $\tau = 0.445$ with the generalization gap at width 384 and $\tau = 0.407$ at width 768, outperforming every baseline, and that its strongest per-axis signal comes from the model-type axis ($\tau = 0.714$). They also claim that minimizing SRR with a stop-gradient at the last layer improves CIFAR-10/100 accuracy by 0.03 to 1.13 points. The paper's secondary discovery is that the canonical MSSA update with a skip connection, equation (8), secretly maximizes the compression term $R_c$ instead of minimizing it; this is traced to the Taylor approximation in equation (7), whose discarded first-order term is what would make descent actually descend. The authors treat these findings as evidence that SRR is a meaningful principle for designing and diagnosing unrolled Transformer-like models, while explicitly limiting the conclusion to the CRATE family.

Load-bearing premise

The headline correlation was measured on models with their normalization layers switched off during evaluation, and the paper gives no evidence that switching them off keeps the models in the same order; if it does not, the correlation is an artifact of how the measure was computed rather than a property of the trained models.

Editorial extensions

If this is right

  • If SRR ranks generalization as reported, then the layer-averaged SRR value of a trained CRATE-family model is a cheap model-selection signal: pick the variant with lower SRR before committing to deployment.
  • Because adding a stop-gradient last-layer SRR penalty improves CIFAR-10/100 accuracy by 0.03–1.13 points, SRR regularization is a low-cost addition to the cross-entropy objective for unrolled models, requiring only one layer's SRR per step.
  • The negative correlation of path-norm with generalization implies that path-norm-based regularizers are unlikely to improve these Transformer-like models, contrary to their role in other settings.
  • The finding that the MSSA update in CRATE-C actually increases $R_c$ implies that the original CRATE's forward pass is not faithfully implementing its stated compress-then-sparsify optimization, and that the sign-flipped CRATE-N is a more faithful instantiation of the compression step.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Because the model-type axis contributes the largest per-axis correlation ($\tau = 0.714$), the headline $\tau$ may mostly separate architectural variants rather than predict generalization within a fixed architecture; testing SRR across many hyperparameter settings inside a single variant, including held-out settings, would clarify this.
  • If the correlation depends on the LayerNorm-removal protocol, then a simpler measure — such as the norm or coding rate of pre-normalization activations — might reproduce the ranking, which would be a cheaper substitute for full SRR; the paper does not test this.
  • SRR regularization with detached activations is a local, layer-wise objective; extending it to per-layer training or to the Forward-Forward style of credit assignment could make it work for very deep models, but that is an extrapolation beyond the paper's experiments.
  • The paper's own limitation that SRR is undefined for standard Transformers with untied query, key, and value matrices suggests a concrete next test: define an analogous coding-rate penalty on attention outputs with untied projections and check whether the positive correlation to generalization survives outside the CRATE family.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. This paper investigates the Sparse Rate Reduction (SRR) objective in the white-box Transformer-like CRATE architecture. The authors first analyze the MSSA update with a skip connection and argue, via a Taylor-expansion argument and a toy experiment, that the standard CRATE-C implementation actually increases the coding rate Rc rather than decreasing it. They then introduce two variants, CRATE-N and CRATE-T, and study how SRR evolves across layers and training epochs. In the central empirical study, they train 64 CRATE-family models on CIFAR-10 and report Kendall rank correlations between a layer-averaged SRR complexity measure and the generalization gap, finding a positive overall correlation (tau = 0.445 at width 384 and 0.407 at width 768) that outperforms several baseline complexity measures. Finally, they show that adding an SRR-based regularizer to the cross-entropy loss yields small accuracy improvements on CIFAR-10/100.

Significance. If the correlation result is valid, the paper makes a useful contribution by proposing a representation-based complexity measure for an interpretable Transformer-like architecture and by connecting an information-theoretic training objective to generalization. The paper is honest about the scope of its conclusions: the Limitations section explicitly notes that the findings are restricted to the CRATE family and that the analysis is limited in scale. The derivation of the decompression artifact in Section 4.1 is a concrete and testable observation, and the toy experiment plus the Appendix A ablations give a clear demonstration that the MSSA update with softmax increases Rc. The introduction of CRATE-N and CRATE-T as design variants, and the comparison of their layer-wise SRR behavior, are also useful. The central limitation, however, is that the headline correlation is computed under an evaluation protocol that removes LayerNorm from models trained with LayerNorm, and no evidence is provided that this protocol preserves the ranking of models by SRR.

major comments (3)
  1. [Appendix C, Section 5.2] The correlation study computes all complexity measures on a forward pass with LayerNorm removed, although every model was trained with LayerNorm before each operator. Because the SRR measure in eq. (11) contains log-determinant coding-rate terms and an l0 count, both of which are sensitive to the scale and location of activations, removing LayerNorm changes the function computed by the trained model. The paper provides no sensitivity analysis or rank-preservation check for any of the measures, including SRR. Without evidence that model rankings by SRR are stable under this protocol change, the reported tau = 0.445 (and tau = 0.407 at width 768) could be an artifact of the evaluation protocol rather than a property of the learned representations. Please report correlations with LayerNorm included at evaluation time, or an ablation demonstrating that the rank ordering of the 64 models is preserved when LayerNorm is removed.
  2. [Section 5.2, Tables 1 and 5] The claim that SRR has 'predictive power' for generalization is supported only by rank correlations computed on the same 64 in-sample models whose generalization gaps are used for the correlation. No held-out hyperparameter settings are tested, and the width 384 is selected for presentation because the correlations are 'more prominent' there. At width 384, the model-type axis alone gives tau = 0.714 for SRR, which is much larger than the overall tau = 0.445, suggesting that model-type differences dominate the correlation. To support the predictive-power claim, the authors should report an out-of-sample evaluation, such as leave-one-hyperparameter-out analysis, or explicitly condition on the model-type axis and show that SRR remains predictive within each model family.
  3. [Section 6, Eq. (14)] The regularization experiments are presented as evidence that SRR causally improves generalization, but the reported gains are small (0.03 to 1.13 percentage points), the regularization coefficient eta = 0.001 is selected by a grid search, and no comparison is made against regularizing with other complexity measures (e.g., path-norm, sharpness, or a simple activation-norm penalty). Since the stop-gradient and last-layer-only implementation is a specific design choice, it is unclear whether the improvement comes from the SRR objective itself or from any layer-wise activation regularizer. Please provide ablations that compare SRR regularization with alternative complexity-measure regularizers under the same protocol, and report variance over multiple seeds.
minor comments (6)
  1. [Eq. (14)] The symbol lambda is used both for the sparsity weight inside the SRR measure, as in eq. (11), and for the overall regularization coefficient in eq. (14). These are different quantities and should be denoted by different symbols to avoid ambiguity.
  2. [Eq. (13)] The displayed Kendall tau formula does not account for ties in the complexity measure or the generalization gap. The standard Kendall tau-b correction, or a statement that no ties occur, should be included so the reported coefficients are reproducible.
  3. [Figures 1(b), 2, and 3] The figure quality is inconsistent: Figure 1(b) has garbled axis labels, and Figures 2 and 3 contain many overlapping curves that are difficult to read in print. Please reformat these figures and, where possible, add markers or a small multiples layout.
  4. [Section 4.3] The sentence 'in the absence of operation (3)' is confusing because operation (3) is in the same layer as the MSSA operator; the intended meaning appears to be 'when the sparsification step is disabled,' which should be stated explicitly.
  5. [Table 6] The efficient-implementation comparison reports single accuracies without standard deviations or seed information, making it hard to judge whether differences such as 77.61 vs. 77.75 are meaningful. Please report multiple runs or at least state that results are from a single seed.
  6. [Section 5.2] The phrase 'correlation with the generation' in the experimental-results paragraph appears to be a typo for 'generalization' and should be corrected.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: SRR is an independent empirical correlate, not fitted to the generalization gap.

full rationale

The paper's central empirical claim is that the fixed complexity measure µ_SRR defined in Eq. (11) correlates with the generalization gap across 64 trained CRATE-family models. The measure is computed from the model's own coding-rate terms and ℓ0 norms; it is not constructed from, or fitted to, the validation or training loss that defines the generalization gap. The Kendall τ in Eq. (13) is an evaluation statistic, not an optimized objective, and no parameter of µ_SRR is tuned to maximize that correlation (λ is fixed at 0.1 in Section 4.3). The SRR objective and CRATE architecture are imported from prior work [45], which shares no authors with the present paper, so the citation is external evidence rather than a self-citation chain. The Section 4.1 critique of the MSSA operator is a self-contained Taylor-expansion argument and does not presuppose the conclusion about generalization. The main methodological limitations—LayerNorm removal during evaluation (Appendix C) and the in-sample nature of the correlation—are validity or generalizability concerns, not circular reductions, because the measure and the target quantity are not equal by construction. The paper also honestly restricts its conclusion to the CRATE family in the Limitations section. Therefore no circular step meets the required standard of exhibiting an equation-level reduction or a fitted parameter renamed as a prediction.

Assumptions & free parameters 3 free parameters · 6 assumptions · 2 invented entities

All free parameters are stated explicitly and inherited from [45] or chosen by grid search; the unexamined choices are the sparsity weight lambda = 0.1 in the measure and the removal of LayerNorm at evaluation time. The two new model variants are fully specified transformations with falsifiable behavioral claims checked in the paper, not unexplained postulates. No new particles, forces, or conservation laws are introduced.

free parameters (3)
  • Sparsity weight lambda in the SRR measure = 0.1
    Appears in the complexity measure (eq. 11) and in the regularization target; taken from [45] with no sensitivity analysis, so the reported correlation could depend on this value.
  • SRR regularization coefficient eta = 0.001
    Tuned by grid search over {0.0001, 0.001, 0.01, 0.1, 1} on the same CIFAR-10/100 evaluations reported in Table 2; the best value is used for the accuracy claims.
  • Toy-experiment step size alpha and coding-rate scaling gamma = alpha = 1, gamma = 1
    Section 4.1; the decompression artifact is demonstrated only at these values with epsilon^2 chosen so gamma = 1; dependence of the artifact on these constants is not analyzed.
assumptions (6)
  • standard math log(1 + x) >= x - x^2/2 for all x >= 0 (the eigenvalues lambda_i of I + gamma*G satisfy lambda_i >= 1)
    Eq. (5) builds a lower bound for Rc from this inequality; the bound is valid and the subsequent gradients in eqs. (6)-(8) follow from it.
  • standard math Gradient identities for the coding rate and its Taylor terms
    Eqs. (6)-(7) compute the gradient of Rc exactly and then of its first- and second-order Taylor terms; these identities are standard and check out.
  • domain assumption Softmax of the Gram matrix is a valid approximation of the attention weights in the unrolled update
    Eq. (8) substitutes softmax((U_k^T Z)^T (U_k^T Z)) for the exact second-order term; this is inherited from the CRATE derivation in [45] Appendix A.2, and Appendix A Fig. 5 shows the artifact persists with or without softmax.
  • ad hoc to paper LayerNorm can be removed at evaluation time without changing the ranking of models by any complexity measure
    Appendix C states LayerNorm is included before each operator during training but removed when evaluating the complexity measures; all 64 models in the correlation study were trained with LayerNorm, and no sensitivity check is given.
  • domain assumption Models that reach a training cross-entropy loss of 0.01 form a representative sample of the hyperparameter grid
    Section 5.2 stops each run at CE loss 0.01 following [18] and excludes models that never reach it, without reporting how many were dropped.
  • domain assumption d = Kp (width equals product of subspaces and subspace dimension) and N = 196 patch tokens
    Sections 3 and 4.1; the toy and main experiments inherit this ViT/CRATE configuration, which fixes p = 64 for d = 384, K = 6.
invented entities (2)
  • CRATE-N, the negative-sign MSSA update (eq. 9) independent evidence
    purpose: Provide a CRATE variant whose forward pass descends on the compression objective Rc as designed, countering the artifact found in CRATE-C
    Fully specified transformation; the paper trains it on CIFAR-10/100 and measures its layer-wise SRR behavior, so its properties are empirically checkable.
  • CRATE-T, the transposed output projection (eq. 10) independent evidence
    purpose: Close the accuracy gap to the full CRATE without adding learnable parameters and while keeping the shared basis U interpretable
    Fully specified transformation; it is the strongest variant in Tables 2 and 3 and is analyzed in the correlation and regularization studies.

how reviews work

0 comments
Cite this review

Pith. "Pith review of An In-depth Investigation of Sparse Rate Reduction in Transformer-like Models." pith.science (2026). https://pith.science/paper/W7OIS7IE

@misc{pith2026241117182,
  author       = {Pith},
  title        = {Pith review of: An In-depth Investigation of Sparse Rate Reduction in Transformer-like Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/W7OIS7IE}},
  note         = {Machine review of arXiv:2411.17182}
}
read the original abstract

Deep neural networks have long been criticized for being black-box. To unveil the inner workings of modern neural architectures, a recent work \cite{yu2024white} proposed an information-theoretic objective function called Sparse Rate Reduction (SRR) and interpreted its unrolled optimization as a Transformer-like model called Coding Rate Reduction Transformer (CRATE). However, the focus of the study was primarily on the basic implementation, and whether this objective is optimized in practice and its causal relationship to generalization remain elusive. Going beyond this study, we derive different implementations by analyzing layer-wise behaviors of CRATE, both theoretically and empirically. To reveal the predictive power of SRR on generalization, we collect a set of model variants induced by varied implementations and hyperparameters and evaluate SRR as a complexity measure based on its correlation with generalization. Surprisingly, we find out that SRR has a positive correlation coefficient and outperforms other baseline measures, such as path-norm and sharpness-based ones. Furthermore, we show that generalization can be improved using SRR as regularization on benchmark image classification datasets. We hope this paper can shed light on leveraging SRR to design principled models and study their generalization ability.

Figures

Figures reproduced from arXiv: 2411.17182 by the authors.

Figure 1
Figure 1. In a simplified attention-only experiment, MSSA operator with skip connection actually [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Sparse rate reduction measure λ∥Z∥0 + Rc (Z; U) − R(Z) of CRATE and its variants evaluated at different layers and epochs on CIFAR-10. at each layer will increase monotonically across all models, with a rare exception in the last few layers of CRATE-C. These phenomena give us implications for understanding Transformer-like models: the representations of initialized models converge fast in the first few layers and ho… view at source ↗
Figure 3
Figure 3. Sparse rate reduction measure λ∥Z∥0 + Rc (Z; U) − R(Z) of CRATE and its variants evaluated at different layers and epochs on CIFAR-100. complexity measure that belongs to the latter category: µSRR(w; Z) = 1 L X L ℓ=1 µ ℓ SRR(wℓ ; Z ℓ ) = 1 L X L ℓ=1  λ∥Z ℓ ∥0 + R c (Z ℓ ; U ℓ ) − R(Z ℓ )  , (11) where Z ℓ denotes the output at layer ℓ and wℓ contains the parameters including U ℓ and Dℓ . 5.2 Correlation with Gener… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: A scatter plot illustrating the value of SRR measure and generalization gap across CRATE and variants with network width d = 384. We confirm the findings from prior works that some norm-based measures, such as sum/prod of spectral/Frobenius norm of parameters neg￾ative…
Figure 5
Figure 5. Figure 5: (a) Original gradient update, i.e, (6). (b) Update from second-order Taylor expansion, i.e., (7). (c) Update from removing the second-order term from (7). (d) Update from removing the first-order term from (7). (e) Update from further adding softmax, i.e., (8). B Diffe…
Figure 6
Figure 6. Figure 6: A scatter plot illustrating the value of SRR measure and generalization gap across CRATE and variants with network width d = 768 [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

48 extracted references · 27 canonical work pages

  1. [1]

    Repulsive attention: Rethinking multi-head attention as bayesian inference

    Bang An, Jie Lyu, Zhenyi Wang, Chunyuan Li, Changwei Hu, Fei Tan, Ruiyi Zhang, Yifan Hu, and Changyou Chen. Repulsive attention: Rethinking multi-head attention as bayesian inference. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 236–255, 2020

    work page 2020

  2. [2]

    Layer normalization

    Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016

    arXiv 2016

  3. [3]

    Spectrally-normalized margin bounds for neural networks

    Peter L Bartlett, Dylan J Foster, and Matus J Telgarsky. Spectrally-normalized margin bounds for neural networks. Advances in neural information processing systems, 30, 2017

    2017

  4. [4]

    Birth of a transformer: A memory viewpoint

    Alberto Bietti, Vivien Cabannes, Diane Bouchacourt, Herve Jegou, and Leon Bottou. Birth of a transformer: A memory viewpoint. Advances in Neural Information Processing Systems, 36, 2023

    work page 2023

  5. [5]

    Attention approximates sparse distributed memory

    Trenton Bricken and Cengiz Pehlevan. Attention approximates sparse distributed memory. Advances in Neural Information Processing Systems, 34:15301–15315, 2021

    work page 2021

  6. [6]

    Invariant scattering convolution networks

    Joan Bruna and Stéphane Mallat. Invariant scattering convolution networks. IEEE transactions on pattern analysis and machine intelligence, 35(8):1872–1886, 2013. 10

    work page 2013

  7. [7]

    Emerging properties in self-supervised vision transformers

    Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650–9660, 2021

    2021

  8. [8]

    Transformer interpretability beyond attention visualization

    Hila Chefer, Shir Gur, and Lior Wolf. Transformer interpretability beyond attention visualization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 782–791, 2021

    work page 2021

Show all 48 references
  1. [9]

    Randaugment: Practical automated data augmentation with a reduced search space

    Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 702–703, 2020

    2020

  2. [10]

    Analyzing transformers in embedding space

    Guy Dar, Mor Geva, Ankit Gupta, and Jonathan Berant. Analyzing transformers in embedding space. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 16124–16170, 2023

    2023

  3. [11]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Con...

    2020

  4. [12]

    A mathematical framework for transformer circuits

    Nelson Elhage, Neel Nanda, Catherine Olsson, Tom Henighan, Nicholas Joseph, Ben Mann, Amanda Askell, Yuntao Bai, Anna Chen, Tom Conerly, Nova DasSarma, Dawn Drain, Deep Ganguli, Zac Hatfield-Dodds, Danny Hernandez, Andy Jones, Jackson Kernion, Liane Lovitt, Kamal Ndousse, Dari...

  5. [13]

    The emergence of clusters in self-attention dynamics

    Borjan Geshkovski, Cyril Letrouit, Yury Polyanskiy, and Philippe Rigollet. The emergence of clusters in self-attention dynamics. Advances in Neural Information Processing Systems, 36, 2023

    2023

  6. [14]

    Patchscopes: A unifying framework for inspecting hidden representations of language models

    Asma Ghandeharioun, Avi Caciularu, Adam Pearce, Lucas Dixon, and Mor Geva. Patchscopes: A unifying framework for inspecting hidden representations of language models. In Forty-first International Conference on Machine Learning, 2024

    2024

  7. [15]

    Learning fast approximations of sparse coding

    Karol Gregor and Yann LeCun. Learning fast approximations of sparse coding. In Proceedings of the 27th international conference on international conference on machine learning, pages 399–406, 2010

    2010

  8. [16]

    The forward-forward algorithm: Some preliminary investigations

    Geoffrey Hinton. The forward-forward algorithm: Some preliminary investigations. arXiv preprint arXiv:2212.13345, 2022

    2022 arXiv

  9. [17]

    Energy transformer

    Benjamin Hoover, Yuchen Liang, Bao Pham, Rameswar Panda, Hendrik Strobelt, Duen Horng Chau, Mohammed Zaki, and Dmitry Krotov. Energy transformer. Advances in Neural Informa- tion Processing Systems, 36, 2023

    2023

  10. [18]

    Fantastic generalization measures and where to find them

    Yiding Jiang*, Behnam Neyshabur*, Hossein Mobahi, Dilip Krishnan, and Samy Bengio. Fantastic generalization measures and where to find them. In International Conference on Learning Representations, 2020

    2020

  11. [19]

    A new measure of rank correlation

    Maurice G Kendall. A new measure of rank correlation. Biometrika, 30(1/2):81–93, 1938

    1938

  12. [20]

    On large-batch training for deep learning: Generalization gap and sharp minima

    Nitish Shirish Keskar, Dheevatsa Mudigere, Jorge Nocedal, Mikhail Smelyanskiy, and Ping Tak Peter Tang. On large-batch training for deep learning: Generalization gap and sharp minima. In International Conference on Learning Representations, 2016

    2016

  13. [21]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Yoshua Bengio and Yann LeCun, editors, 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015

    2015

  14. [22]

    Tracr: Compiled transformers as a laboratory for interpretability

    David Lindner, János Kramár, Sebastian Farquhar, Matthew Rahtz, Tom McGrath, and Vladimir Mikulik. Tracr: Compiled transformers as a laboratory for interpretability. Advances in Neural Information Processing Systems, 36, 2023. 11

    2023

  15. [23]

    Omnigrok: Grokking beyond algorithmic data

    Ziming Liu, Eric J Michaud, and Max Tegmark. Omnigrok: Grokking beyond algorithmic data. In The Eleventh International Conference on Learning Representations, 2022

    2022

  16. [24]

    Segmentation of multivariate mixed data via lossy data coding and compression

    Yi Ma, Harm Derksen, Wei Hong, and John Wright. Segmentation of multivariate mixed data via lossy data coding and compression. IEEE transactions on pattern analysis and machine intelligence, 29(9):1546–1562, 2007

    2007

  17. [25]

    Pac-bayesian model averaging

    David A McAllester. Pac-bayesian model averaging. In Proceedings of the twelfth annual conference on Computational learning theory, pages 164–170, 1999

    1999

  18. [26]

    Universal hopfield networks: A general framework for single-shot associative memory models

    Beren Millidge, Tommaso Salvatori, Yuhang Song, Thomas Lukasiewicz, and Rafal Bogacz. Universal hopfield networks: A general framework for single-shot associative memory models. In International Conference on Machine Learning, pages 15561–15583. PMLR, 2022

    2022

  19. [27]

    Algorithm unrolling: Interpretable, efficient deep learning for signal and image processing

    Vishal Monga, Yuelong Li, and Yonina C Eldar. Algorithm unrolling: Interpretable, efficient deep learning for signal and image processing. IEEE Signal Processing Magazine, 38(2):18–44, 2021

    2021

  20. [28]

    Progress measures for grokking via mechanistic interpretability

    Neel Nanda, Lawrence Chan, Tom Lieberum, Jess Smith, and Jacob Steinhardt. Progress measures for grokking via mechanistic interpretability. InThe Eleventh International Conference on Learning Representations, 2023

    2023

  21. [29]

    Exploring generalization in deep learning

    Behnam Neyshabur, Srinadh Bhojanapalli, David McAllester, and Nati Srebro. Exploring generalization in deep learning. Advances in neural information processing systems, 30, 2017

    2017

  22. [30]

    A PAC-bayesian approach to spectrally-normalized margin bounds for neural networks

    Behnam Neyshabur, Srinadh Bhojanapalli, and Nathan Srebro. A PAC-bayesian approach to spectrally-normalized margin bounds for neural networks. In International Conference on Learning Representations, 2018

    2018

  23. [31]

    Path-sgd: Path-normalized optimization in deep neural networks

    Behnam Neyshabur, Russ R Salakhutdinov, and Nati Srebro. Path-sgd: Path-normalized optimization in deep neural networks. Advances in neural information processing systems, 28, 2015

    2015

  24. [32]

    Norm-based capacity control in neural networks

    Behnam Neyshabur, Ryota Tomioka, and Nathan Srebro. Norm-based capacity control in neural networks. In Conference on learning theory, pages 1376–1401. PMLR, 2015

    2015

  25. [33]

    Theoretical foundations of deep learning via sparse representations: A multilayer sparse model and its connection to convolutional neural networks

    Vardan Papyan, Yaniv Romano, Jeremias Sulam, and Michael Elad. Theoretical foundations of deep learning via sparse representations: A multilayer sparse model and its connection to convolutional neural networks. IEEE Signal Processing Magazine, 35(4):72–89, 2018

    2018

  26. [34]

    Grokking: Gen- eralization beyond overfitting on small algorithmic datasets

    Alethea Power, Yuri Burda, Harri Edwards, Igor Babuschkin, and Vedant Misra. Grokking: Gen- eralization beyond overfitting on small algorithmic datasets. arXiv preprint arXiv:2201.02177, 2022

    2022 arXiv

  27. [35]

    Hopfield networks is all you need

    Hubert Ramsauer, Bernhard Schäfl, Johannes Lehner, Philipp Seidl, Michael Widrich, Lukas Gruber, Markus Holzleitner, Thomas Adler, David Kreil, Michael K Kopp, Günter Klambauer, Johannes Brandstetter, and Sepp Hochreiter. Hopfield networks is all you need. In International Con...

    2021

  28. [36]

    Unraveling attention via convex duality: Analysis and interpretations of vision transformers

    Arda Sahiner, Tolga Ergen, Batu Ozturkler, John Pauly, Morteza Mardani, and Mert Pilanci. Unraveling attention via convex duality: Analysis and interpretations of vision transformers. In International Conference on Machine Learning, pages 19050–19088. PMLR, 2022

    2022

  29. [37]

    Biological learning in key-value memory networks

    Danil Tyulmankov, Ching Fang, Annapurna Vadaparty, and Guangyu Robert Yang. Biological learning in key-value memory networks. Advances in Neural Information Processing Systems, 34:22247–22258, 2021

    2021

  30. [38]

    On the uniform convergence of relative frequencies of events to their probabilities

    Vladimir N Vapnik and A Ya Chervonenkis. On the uniform convergence of relative frequencies of events to their probabilities. In Measures of complexity: festschrift for alexey chervonenkis, pages 11–30. Springer, 2015

    2015

  31. [39]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 12

    2017

  32. [40]

    Interpretability in the wild: a circuit for indirect object identification in GPT-2 small

    Kevin Ro Wang, Alexandre Variengien, Arthur Conmy, Buck Shlegeris, and Jacob Steinhardt. Interpretability in the wild: a circuit for indirect object identification in GPT-2 small. In The Eleventh International Conference on Learning Representations, 2023

    2023

  33. [41]

    Thinking like transformers

    Gail Weiss, Yoav Goldberg, and Eran Yahav. Thinking like transformers. In International Conference on Machine Learning, pages 11080–11090. PMLR, 2021

    2021

  34. [42]

    Graph neural networks inspired by classical iterative algorithms

    Yongyi Yang, Tang Liu, Yangkun Wang, Jinjing Zhou, Quan Gan, Zhewei Wei, Zheng Zhang, Zengfeng Huang, and David Wipf. Graph neural networks inspired by classical iterative algorithms. In International Conference on Machine Learning, pages 11773–11783. PMLR, 2021

    2021

  35. [43]

    Transformers from an optimization perspective

    Yongyi Yang, David P Wipf, et al. Transformers from an optimization perspective. Advances in Neural Information Processing Systems, 35:36958–36971, 2022

    2022

  36. [44]

    Attentionviz: A global view of transformer attention

    Catherine Yeh, Yida Chen, Aoyu Wu, Cynthia Chen, Fernanda Viégas, and Martin Wattenberg. Attentionviz: A global view of transformer attention. IEEE Transactions on Visualization and Computer Graphics, 2023

    2023

  37. [45]

    White-box transformers via sparse rate reduction

    Yaodong Yu, Sam Buchanan, Druv Pai, Tianzhe Chu, Ziyang Wu, Shengbang Tong, Benjamin Haeffele, and Yi Ma. White-box transformers via sparse rate reduction. Advances in Neural Information Processing Systems, 36, 2023

    2023

  38. [46]

    Learning diverse and discriminative representations via the principle of maximal coding rate reduction

    Yaodong Yu, Kwan Ho Ryan Chan, Chong You, Chaobing Song, and Yi Ma. Learning diverse and discriminative representations via the principle of maximal coding rate reduction. Advances in Neural Information Processing Systems, 33:9422–9434, 2020

    2020

  39. [47]

    Unveiling transformers with lego: a synthetic reasoning task

    Yi Zhang, Arturs Backurs, Sébastien Bubeck, Ronen Eldan, Suriya Gunasekar, and Tal Wagner. Unveiling transformers with lego: a synthetic reasoning task. arXiv preprint arXiv:2206.04301, 2022. 13 A Complete Demonstrations of the Pitfalls To give a clearer picture of how approxi...

    2022 arXiv

  40. [2021]

    https://transformer-circuits.pub/2021/framework/index.html

    2021

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.