Pith. sign in

REVIEW 4 major objections 6 minor 2 cited by

Same Compression Principle, Different Geometry: Rate-Distortion Signatures Dissociate Biological and Artificial Visual Systems

T0 review · 4 major / 6 minor · reviewed 2026-08-02 · deepseek-v4-flash

Pith's one-line read Both biological and artificial visual systems obey a common lossy-compression principle, yet rate–distortion geometry separates humans from deep networks: humans trade fidelity smoothly, networks in steep, brittle steps.

desk verdict Useful descriptive study of RD-style confusion geometry in 18 vision models, but the central dissociation is a reparameterization of a fitted cost matrix, and the abstract overclaims an internal-representation link the paper does not contain. read the letter →

arxiv 2603.01568 v2 pith:STFX4QEJ submitted 2026-03-02 cs.LG cs.CVcs.ITmath.ITq-bio.NC

classification cs.LGcs.CVcs.ITmath.ITq-bio.NC
keywords rate-distortiontheoryconfusionmatricesefficientcodingvisualgeneralizationhumanvsmachinevisionrobustnessevaluationrepresentationgeometrycompressiontrade-off
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

Visual systems, human or artificial, can be viewed as communication channels that compress noisy images into labels. The paper infers, from each system's confusion matrix, a rate–distortion trade-off curve and summarizes it with three numbers: slope, curvature, and area under the curve. It claims that all systems follow the same lossy-compression principle, but that humans occupy a low-slope, low-curvature region—smooth, flexible degradation—while deep network families sit in steeper, more brittle regions even when accuracy matches. Training for robustness shifts models along different axes of this space, so 'human-likeness' is not a single scale. The paper also argues that these behaviorally inferred compression signatures track internal representational geometry, making the diagnostic informative about internal structure without looking inside the network.

What carries the argument

The load-bearing object is the behavioral rate–distortion curve R(D), built from a row-normalized K×K confusion matrix treated as an effective channel p(y|x). The paper infers a K×K cost matrix ρ by maximum a posteriori fitting under a structured prior, then scales ρ by an inverse-temperature λ and iterates fixed-point updates to trace the frontier. From the traced curve it extracts median slope β (typical marginal information cost), the variance of local slopes κ (how abruptly the system switches between regimes), and area AUC (overall efficiency). These three numbers are the signature that organizes humans and models on a shared trade-off surface.

What would settle it

Compute RD signatures for the same humans and models under a fixed symmetric 0–1 cost matrix instead of the MAP-inferred ρ. If humans and models then overlap in β/κ, or if ρ is non-unique across random restarts of the MAP fit, the paper's central dissociation fails. A second check: if a network with decorrelated internal representations produces identical behavioral RD signatures, the claim that signatures track internal representational geometry is contradicted.

Watch

Extended reading notes

Core claim

The central claim is that confusion statistics alone are enough to reconstruct each system's effective rate–distortion frontier, and that the shape of this frontier is a stable signature of the system's compression strategy. On the paper's own evidence, humans show gradual, exponential-like generalization gradients (low slope, low curvature); convolutional baselines, local-feature models, shape-biased models, and several self-/semi-supervised families show displaced, steeper frontiers; and robustness-training regimes can approach humans on one axis while overshooting on another. The paper further asserts that the fitted cost geometry correlates with internal representational dissimilarity, s

Load-bearing premise

The dissociation is load-bearing on the assumption that the cost matrix ρ is identifiable from off-diagonal confusion counts and that the structured prior does not itself create the smooth human geometry; if that fails, the human/model split is an artifact of fitting.

Editorial extensions

If this is right

  • Rate–distortion geometry separates model families even when accuracy does not: some models match human accuracy yet have significantly steeper slope and higher curvature.
  • Robustness training moves models directionally: distortion-trained models shift toward human geometry at a cost in accuracy and efficiency, while multi-corruption and single-distortion training improve accuracy and efficiency but overshoot past humans on curvature.
  • Because behavioral RD signatures track internal representational dissimilarity, a researcher can probe representation-level compression structure from behavior alone, without accessing activations.
  • RD signatures offer a complementary model-selection criterion: when two systems match in accuracy, lower curvature indicates smoother degradation under perturbation, and higher AUC indicates better overall efficiency in the tested regime.
  • The same pipeline applies to adversarial and certified-robust networks, distinguishing robustness strategies that achieve similar accuracy but different degradation profiles.

Reading between the lines

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

  • A testable extension: re-run the entire analysis with fixed 0–1 cost instead of inferred ρ; if the human/model dissociation vanishes, the dissociation comes from the fitting procedure, not the systems.
  • A predicted consequence for training research: models trained with smooth, curriculum-like augmentation should drift toward the human low-κ region, while adversarial training should push κ up; this is measurable with the same pipeline.
  • For neuroscience, the same effective-channel logic applied to population activity would predict that biological tuning smoothness yields near-zero κ, providing a direct bridge from behavior to coding efficiency.
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

4 major / 6 minor

Summary. The paper proposes a rate-distortion (RD) framework for comparing human and deep-network visual systems using confusion matrices. It infers a per-experiment cost matrix rho via MAP estimation, derives a behavioral RD curve from the optimal channel q_lambda(y|x) ∝ q(y) exp(-lambda rho(x,y)), and summarizes each system with three signatures: slope (beta), curvature (kappa), and AUC. Applying this to human psychophysics and 18 vision models across 12 perturbation families, the paper finds that humans occupy a low-slope, low-curvature regime while deep networks are steeper and more brittle, with robustness training shifting models along different geometric axes. The authors claim this provides a compact, model-agnostic diagnostic beyond accuracy-based robustness metrics.

Significance. If the methodological concerns are resolved, the framework would be a useful complement to standard robustness evaluation: it is model-agnostic, works on matched human and machine data, and the statistical analysis is careful (block-aware paired Wilcoxon tests, FDR control, fixed-effects regressions showing that RD geometry captures variance beyond accuracy). The use of two large public benchmarks (GEN and ModelZoo) is a strength. However, the central empirical dissociation currently rests on an unvalidated fitting procedure and an unsupported abstract claim; the paper's own limitations section acknowledges that key sensitivity analyses are left for future work. These issues are addressable in revision, but they are load-bearing for the main contribution.

major comments (4)
  1. [§2.2, §4, Limitations] The central dissociation is computed from the MAP-inferred cost matrix ρ, which is fitted to the same confusion matrices that define each system. Because ρ is identifiable only up to row-wise constants and global scaling, and because the structured prior (whose strength is never reported) can impose smoothness that later appears as low κ for humans, the reported human/model separation may be an artifact of the fitting procedure. The paper itself defers 'cost inference vs. fixed distortion (e.g., 0–1)' and prior-strength checks to future work, but these are load-bearing for the main claim. Please report the prior hyperparameters, perform an identifiability/posterior-predictive analysis, and recompute β/κ/AUC under a fixed 0-1 cost and under multiple prior strengths. Without these, the claim that humans and deep networks occupy systematically different regions of RD space is not establishe
  2. [§2.3, Fig. 2, Table 2] The sign convention for β is inconsistent. §2.3 defines β as the median of finite-difference slopes s_k = (R_{k+1}-R_k)/(D_{k+1}-D_k), which are negative for a decreasing RD curve. Yet the text describes models as having 'steeper slopes' and Table 2 reports positive Δβ for models versus humans, while Fig. 2 uses log10 |β|. This ambiguity affects every quantitative comparison in the paper. Please define whether β is reported as the signed slope or its absolute value, apply the same convention throughout, and re-examine the 'steeper' conclusions under both conventions.
  3. [Abstract (arXiv) vs. full text, §5] The arXiv abstract and title claim that 'behavioral RD signatures track internal representational geometry, evidenced by ... correlation with internal representational dissimilarity across all models.' No such analysis appears anywhere in the provided full text; the Discussion explicitly defers 'relat[ing] behavioral RD geometry to internal representational structure' to future work. This is a direct mismatch between the advertised contribution and the manuscript. Either add the analysis or revise the abstract/title to describe what is actually demonstrated.
  4. [§4.1, Appendix B] The fit diagnostic RMSE≈0.078 (and Table 1) is presented as evidence that an effective RDT channel explains confusion structure, but no null model or baseline is given. A K×K cost matrix with a structured prior plus a λ sweep has many degrees of freedom; low RMSE does not validate the exponential-family form. Please compare against simpler baselines (e.g., independent errors, 0-1 cost, or a channel without cost structure) and report cross-validated or parameter-count-adjusted measures.
minor comments (6)
  1. [Title] The arXiv title 'Same Compression Principle, Different Geometry...' differs from the full-text title 'Rate–Distortion Signatures of Generalization and Information Trade-off...'. Please align them.
  2. [§2.2] The Laplace approximation for posterior uncertainty is mentioned but never used. Either use it or remove the claim.
  3. [§2.3] The λ grid is described only as 'log-spaced'; for reproducibility, specify the range, number of points, and how endpoints/finite differences are handled.
  4. [§4.1] The text reports a pooled RMSE of 0.078, but Table 1 does not include this pooled value. Clarify how it is computed.
  5. [Appendix A.2, Fig. 3] The legend lists 'VGG-19' twice; the second occurrence should likely be 'Baseline CNNs' or similar.
  6. [References] Jakob & Gershman (2023a) and (2023b) appear to be the same work; unify the citations.

Circularity Check

2 steps flagged · score 6.0 of 10

RD geometry is a re-parameterization of the MAP-fitted cost matrix; in-sample 'predictions' double as validation.

  1. fitted input called prediction [§4.1 / Fig. 1a (with §2.2)]
    "RDT successfully captures the dominant structure of confusions and generalization gradients with consistently low error RMSE= 0.078 ; see Appendix B for all values). Figure 1a shows empirical generalization probabilities G(x,ˆx) as a function of inferred error cost ρ, together with the corresponding RDT-predicted curves."

    The 'RDT-predicted curves' are generated by Eq. 3 from the MAP-inferred cost matrix ρ, which was fit to the very confusion matrices being compared (the paper states ρ is inferred 'from empirical confusions' and the implied channel is evaluated by comparing it to 'the empirical behavior'). Thus the low RMSE is an in-sample goodness-of-fit, not an out-of-sample prediction, and cannot independently confirm that systems 'follow a common lossy-compression principle'; the flexible per-system ρ makes the RD channel able to approximate arbitrary confusion structure.

  2. self definitional [§2.2–§2.3 (Eq. 3 and signature definitions)]
    "Given the inferred cost matrix ρ, we trace a discrete approximation to the rate–distortion frontier by scaling ρ with an inverse-temperature parameter λ>0 and computing the corresponding optimal channel: qλ(y|x)∝q λ(y) exp{−λ ρ(x, y)}. ... From each estimated rate–distortion frontier {(Dk, Rk)}K k=1, we extract compact summary statistics ... Slope (β). ... β≡median k sk ... Curvature proxy (κ). ... κ≡Var k(sk)."

    β, κ, and AUC are defined as summaries of the RD frontier {(Dk,Rk)} that is computed by Eq. 3 from the inferred ρ. Since ρ is MAP-fitted to the same empirical channel C that defines each system, the signatures are re-parameterizations of the fitted cost matrix rather than independent measurements of an information–fidelity trade-off. The human/model dissociation in RD space is therefore a property of the fitted ρ under the structured prior; identifiability and prior strength are not reported, and the Limitations lists 'cost inference vs. fixed distortion (e.g., 0–1)' as future sensitivity analysis. Without a fixed-cost baseline, the dissociation could be an artifact of the fitting procedure rather than a system property.

full rationale

The paper is self-consistent and does not rely on self-citation; its citations to Sims (2018) are external and appropriate. However, two load-bearing reductions make the central human/model RD dissociation partly circular. First, the cost matrix ρ is MAP-fitted to each system's confusion matrix (§2.2), and the RD curve is then traced from that same ρ (Eq. 3); the reported β, κ, and AUC are summary statistics of that curve (§2.3), so the empirical channel enters only through the fitted ρ. Second, the 'RDT-predicted curves' used for validation in §4.1/Fig. 1a are in-sample reconstructions from the fitted ρ, so low RMSE is a goodness-of-fit, not an independent prediction. The abstract's claim that behavioral RD signatures track internal representational geometry is not supported by any analysis in the provided text—the Discussion explicitly defers relating behavioral RD geometry to internal structure to future work—but that is an unsupported claim rather than a circular step. Because the flexible per-system ρ makes the RD model capable of fitting broad confusion structure, the 'common compression principle' is partly assumed by construction; nevertheless, the observed separations across many models and perturbation families are nontrivial empirical regularities, justifying a score of 6 rather than 8–10.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The paper's central quantities β, κ, AUC are not directly measured; they are functions of the fitted cost matrix ρ plus hand-chosen hyperparameters (prior, smoothing, λ grid). No new physical entities are introduced, but the effective channel itself is a modeling construct.

free parameters (4)
  • Cost matrix ρ (K×K per experiment)
    Fitted via MAP from empirical confusions; all RD signatures are functions of it.
  • Structured prior hyperparameters for ρ
    Regularization strengths for symmetric/asymmetric/diagonal components chosen by hand; directly control smoothness of inferred geometry and hence κ.
  • Laplace smoothing pseudocount α = 0.5
    Used in severity plot (§A.2) to stabilize log-probabilities; affects severity β values, not primary signatures.
  • λ grid (log-spaced)
    Range and density chosen by hand; determines traced RD interval and AUC.
assumptions (4)
  • standard math Blahut-Arimoto fixed-point iterations converge to the rate-distortion function for a given distortion matrix ρ.
    Invoked in §2.2 to trace (D(λ), R(λ)).
  • domain assumption Row-normalized confusion counts approximate the true conditional response distribution p(y|x).
    Eq. 2 in §2.1; limited human trial counts and deterministic model labels make this approximation noisy.
  • ad hoc to paper Systems behave as if optimal for some latent cost ρ, so the fitted exponential-family channel q_λ(y|x) ∝ q(y) exp(−λρ(x,y)) is a valid behavioral model.
    Eq. 3 in §2.2; this form is imposed, not derived from the data, and its fit RMSE ≈0.078 is not compared to a null model.
  • ad hoc to paper The MAP estimate of ρ is identifiable from off-diagonal confusion probabilities under the structured prior.
    Unstated; all signatures depend on this identifiability, and the paper defers sensitivity analysis to future work.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Same Compression Principle, Different Geometry: Rate-Distortion Signatures Dissociate Biological and Artificial Visual Systems." pith.science (2026). https://pith.science/paper/STFX4QEJ

@misc{pith2026260301568,
  author       = {Pith},
  title        = {Pith review of: Same Compression Principle, Different Geometry: Rate-Distortion Signatures Dissociate Biological and Artificial Visual Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/STFX4QEJ}},
  note         = {Machine review of arXiv:2603.01568}
}
read the original abstract

Efficient coding theory predicts that biological perceptual systems compress sensory input optimally under resource constraints, with the systematic structure of errors reflecting the geometry of that compression. Here we operationalize this principle using rate-distortion theory (RDT) to characterize how any system - biological or artificial - trades representational fidelity for informational efficiency. Treating stimulus-response behavior as an effective communication channel, we infer rate-distortion (RD) frontiers directly from confusion matrices and summarize each system with three geometric signatures: slope (beta), curvature (kappa), and area under the RD curve (AUC), capturing the marginal cost, abruptness, and overall efficiency of the accuracy-compression trade-off respectively. Applying this framework to human psychophysical data and 18 deep vision models across 12 families of controlled image perturbations at graded severities, we find that both biological and artificial systems follow a common lossy-compression principle but occupy systematically different regions of RD space. Humans exhibit smooth, flexible trade-offs characteristic of near-optimal efficient coding, while deep networks operate in steeper, more brittle regimes even at matched accuracy, with geometry dissociable from performance across training regimes. Critically, behavioral RD signatures track internal representational geometry, evidenced by the behaviorally inferred compression structure correlating with internal representational dissimilarity across all models. These results establish RD geometry as a compact diagnostic of perceptual compression strategy that recovers mechanistically interpretable structure in internal representations from behavioral input alone and extends naturally to the direct characterization of compression geometry in neural population activity.

Figures

Figures reproduced from arXiv: 2603.01568 by the authors.

Figure 1
Figure 1. Generalization curves and goodness-of-fit of rate–distortion predictions. (a) Empirical generalization G(x, xˆ) (points) plotted against inferred cost of error ρ for humans and selected model families. Red curves show the corresponding exponential/RD-style prediction. Insets show observed vs. predicted off-diagonal confusion probabilities (RDT channel prediction), with RMSE reported per family. (b) Summary of expone… view at source ↗
Figure 2
Figure 2. Rate–distortion signatures across perturbation experiments and system families. (a) For each experiment/condition (x-axis; abbreviations as shown), points show systems’ RD summaries: normalized rate–distortion efficiency (AUC; left), log10 |β| (middle), and log10 |κ| (right), grouped by family/regime (legend). Values are computed per experiment×condition block and visualized to highlight between-family structure acr… view at source ↗
Figure 3
Figure 3. shows that humans exhibit a gradual steepening of β with increasing noise. This means that as the visual input becomes less reliable, the effective channel increases the marginal information cost of accuracy in a relatively smooth manner. By contrast, baseline CNNs show a markedly sharper increase in |β| across the same corruption range, consistent with a more abrupt transition into a high-compression regime. Notabl… view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Directional Confusions Reveal Divergent Inductive Biases Through Rate-Distortion Geometry in Human and Machine Vision

    cs.CV 2026-04 unverdicted novelty 7.0 of 10

    Humans show broad weak directional confusions while DNNs show sparse strong collapses; these structures shift rate-distortion geometry differently and reveal divergent inductive biases.

  2. Directional Confusions Reveal Divergent Inductive Biases Through Rate-Distortion Geometry in Human and Machine Vision

    cs.CV 2026-04 unverdicted novelty 6.0 of 10

    Directional confusions in human and machine vision produce distinct rate-distortion geometric signatures that reveal divergent inductive biases invisible to accuracy metrics alone.

Reference graph

Works this paper leans on

50 extracted references · 9 linked inside Pith · cited by 1 Pith paper

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    A., Fischer, I., Dillon, J

    Alemi, A. A., Fischer, I., Dillon, J. V., and Murphy, K. Deep variational information bottleneck. arXiv preprint arXiv:1612.00410, 2016

  3. [3]

    An algorithm for computing the capacity of arbitrary discrete memoryless channels

    Arimoto, S. An algorithm for computing the capacity of arbitrary discrete memoryless channels. IEEE Transactions on Information Theory, 18 0 (1): 0 14--20, 1972

  4. [4]

    Atick, J. J. and Redlich, A. N. What does the retina know about natural scenes? Neural computation, 4 0 (2): 0 196--210, 1992

  5. [5]

    Some informational aspects of visual perception

    Attneave, F. Some informational aspects of visual perception. Psychological review, 61 0 (3): 0 183, 1954

  6. [6]

    Barlow, H. B. et al. Possible principles underlying the transformation of sensory messages. Sensory communication, 1 0 (01): 0 217--233, 1961

  7. [7]

    Computation of channel capacity and rate-distortion functions

    Blahut, R. Computation of channel capacity and rate-distortion functions. IEEE Transactions on Information Theory, 18 0 (4): 0 460--473, 1972. doi:10.1109/TIT.1972.1054855

  8. [8]

    and Bethge, M

    Brendel, W. and Bethge, M. Approximating cnns with bag-of-local-features models works surprisingly well on imagenet. arXiv preprint arXiv:1904.00760, 2019

Show all 50 references
  1. [9]

    and Geisler, W

    Burge, J. and Geisler, W. S. Optimal disparity estimation in natural stereo images. Journal of vision, 14 0 (2): 0 1--1, 2014

  2. [10]

    A simple framework for contrastive learning of visual representations

    Chen, T., Kornblith, S., Norouzi, M., and Hinton, G. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pp.\ 1597--1607. PmLR, 2020 a

  3. [11]

    Improved baselines with momentum contrastive learning

    Chen, X., Fan, H., Girshick, R., and He, K. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020 b

  4. [12]

    Cover, T. M. and Thomas, J. A. Elements of Information Theory. Wiley, 2nd edition, 2006

  5. [13]

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

    Dosovitskiy, A. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020

  6. [14]

    A systematic review of robustness in deep learning for computer vision: Mind the gap? arXiv preprint arXiv:2112.00639, 2021

    Drenkow, N., Sani, N., Shpitser, I., and Unberath, M. A systematic review of robustness in deep learning for computer vision: Mind the gap? arXiv preprint arXiv:2112.00639, 2021

  7. [15]

    A., and Brendel, W

    Geirhos, R., Rubisch, P., Michaelis, C., Bethge, M., Wichmann, F. A., and Brendel, W. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. In International conference on learning representations, 2018

  8. [16]

    A., and Brendel, W

    Geirhos, R., Narayanappa, K., Mitzkus, B., Thieringer, T., Bethge, M., Wichmann, F. A., and Brendel, W. Partial success in closing the gap between human and machine vision. Advances in Neural Information Processing Systems, 34: 0 23885--23899, 2021

  9. [17]

    Geisler, W. S. Visual perception and the statistical properties of natural scenes. Annu. Rev. Psychol., 59 0 (1): 0 167--192, 2008

  10. [18]

    Adversarial examples are a natural consequence of test error in noise

    Gilmer, J., Ford, N., Carlini, N., and Cubuk, E. Adversarial examples are a natural consequence of test error in noise. In International Conference on Machine Learning, pp.\ 2280--2289. PMLR, 2019

  11. [19]

    Gr \"u nwald, P. D. The minimum description length principle. MIT press, 2007

  12. [20]

    Deep residual learning for image recognition

    He, K., Zhang, X., Ren, S., and Sun, J. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.\ 770--778, 2016

  13. [21]

    and Dietterich, T

    Hendrycks, D. and Dietterich, T. Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261, 2019

  14. [22]

    Using pre-training can improve model robustness and uncertainty

    Hendrycks, D., Lee, K., and Mazeika, M. Using pre-training can improve model robustness and uncertainty. In International conference on machine learning, pp.\ 2712--2721. PMLR, 2019

  15. [23]

    Hern \'a ndez-Garc \' a, A., K \""o nig, P., and Kietzmann, T. C. Learning robust visual representations using data augmentation invariance. arXiv preprint arXiv:1906.04547, 2019

  16. [24]

    Adversarial examples are not bugs, they are features

    Ilyas, A., Santurkar, S., Tsipras, D., Engstrom, L., Tran, B., and Madry, A. Adversarial examples are not bugs, they are features. Advances in neural information processing systems, 32, 2019

  17. [25]

    Jakob, A. M. and Gershman, S. J. Rate-distortion theory of neural coding and its implications for working memory. Elife, 12: 0 e79450, 2023 a

  18. [26]

    Jakob, A. M. and Gershman, S. J. Rate-distortion theory of neural coding and its implications for working memory. Elife, 12: 0 e79450, 2023 b

  19. [27]

    Big transfer (bit): General visual representation learning

    Kolesnikov, A., Beyer, L., Zhai, X., Puigcerver, J., Yung, J., Gelly, S., and Houlsby, N. Big transfer (bit): General visual representation learning. In European conference on computer vision, pp.\ 491--507. Springer, 2020

  20. [28]

    M., and Kietzmann, T

    Lu, Z., Thorat, S., Cichy, R. M., and Kietzmann, T. C. Adopting a human developmental visual diet yields robust, shape-based ai vision. arXiv preprint arXiv:2507.03168, 2025

  21. [29]

    Exploring the limits of weakly supervised pretraining

    Mahajan, D., Girshick, R., Ramanathan, V., He, K., Paluri, M., Li, Y., Bharambe, A., and Van Der Maaten, L. Exploring the limits of weakly supervised pretraining. In Proceedings of the European conference on computer vision (ECCV), pp.\ 181--196, 2018

  22. [30]

    McAllester, D. A. Pac-bayesian model averaging. In Proceedings of the twelfth annual conference on Computational learning theory, pp.\ 164--170, 1999

  23. [31]

    Nosofsky, R. M. Attention, similarity, and the identification--categorization relationship. Journal of experimental psychology: General, 115 0 (1): 0 39, 1986

  24. [32]

    Ortega, P. A. and Braun, D. A. Thermodynamics as a theory of decision-making with information-processing costs. Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences, 469 0 (2153): 0 20120683, 2013

  25. [33]

    W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al

    Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pp.\ 8748--8763. PmLR, 2021

  26. [34]

    Modeling by shortest data description

    Rissanen, J. Modeling by shortest data description. Automatica, 14 0 (5): 0 465--471, 1978

  27. [35]

    Shannon, C. E. et al. Coding theorems for a discrete source with a fidelity criterion. IRE Nat. Conv. Rec, 4 0 (142-163): 0 1, 1959

  28. [36]

    Shepard, R. N. Toward a universal law of generalization for psychological science. Science, 237 0 (4820): 0 1317--1323, 1987

  29. [37]

    E., Torr, P., and Sanyal, A

    Shi, Y., Daunhawer, I., Vogt, J. E., Torr, P., and Sanyal, A. How robust are pre-trained models to distribution shift? In ICML 2022: Workshop on Spurious Correlations, Invariance and Stability, 2022

  30. [38]

    Simoncelli, E. P. and Olshausen, B. A. Natural image statistics and neural representation. Annual review of neuroscience, 24 0 (1): 0 1193--1216, 2001

  31. [39]

    and Zisserman, A

    Simonyan, K. and Zisserman, A. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014

  32. [40]

    Sims, C. R. Efficient coding explains the universal law of generalization in human perception. Science, 360 0 (6389): 0 652--656, 2018

  33. [41]

    P., Mahajan, D., Girshick, R., Doll \'a r, P., and Van Der Maaten, L

    Singh, M., Gustafson, L., Adcock, A., de Freitas Reis, V., Gedik, B., Kosaraju, R. P., Mahajan, D., Girshick, R., Doll \'a r, P., and Van Der Maaten, L. Revisiting weakly supervised pre-training of visual perception models. In Proceedings of the IEEE/CVF Conference on Computer...

  34. [42]

    Going deeper with convolutions

    Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., Erhan, D., Vanhoucke, V., and Rabinovich, A. Going deeper with convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.\ 1--9, 2015

  35. [43]

    Measuring robustness to natural distribution shifts in image classification

    Taori, R., Dave, A., Shankar, V., Carlini, N., Recht, B., and Schmidt, L. Measuring robustness to natural distribution shifts in image classification. Advances in Neural Information Processing Systems, 33: 0 18583--18599, 2020

  36. [44]

    Tenenbaum, J. B. and Griffiths, T. L. Generalization, similarity, and bayesian inference. Behavioral and brain sciences, 24 0 (4): 0 629--640, 2001

  37. [45]

    What makes for good views for contrastive learning? Advances in neural information processing systems, 33: 0 6827--6839, 2020

    Tian, Y., Sun, C., Poole, B., Krishnan, D., Schmid, C., and Isola, P. What makes for good views for contrastive learning? Advances in neural information processing systems, 33: 0 6827--6839, 2020

  38. [46]

    C., and Bialek, W

    Tishby, N., Pereira, F. C., and Bialek, W. The information bottleneck method. arXiv preprint physics/0004057, 2000

  39. [47]

    Optimality principles in sensorimotor control

    Todorov, E. Optimality principles in sensorimotor control. Nature neuroscience, 7 0 (9): 0 907--915, 2004

  40. [48]

    Wichmann, F. A. and Geirhos, R. Are deep neural networks adequate behavioral models of human visual perception? Annual review of vision science, 9 0 (1): 0 501--524, 2023

  41. [49]

    Xie, Q., Luong, M.-T., Hovy, E., and Le, Q. V. Self-training with noisy student improves imagenet classification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 10687--10698, 2020

  42. [50]

    and Raginsky, M

    Xu, A. and Raginsky, M. Information-theoretic analysis of generalization capability of learning algorithms. Advances in neural information processing systems, 30, 2017

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.