Pith. sign in

REVIEW 4 major objections 5 minor 59 references

BIRD: Behavior Induction via Representation-structure Distillation

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Aligned behavior can be transferred across models that differ in architecture, task, and training data by distilling the geometry of their internal representations.

desk verdict The empirical study is useful, but Equation 3 defines CKA incorrectly, so the method as written is not matching representational structure; the paper needs that fixed before its central claim is credible. read the letter →

arxiv 2505.23933 v1 pith:VWFSNC7R submitted 2025-05-29 cs.LG cs.AI

Galen Pogoncheff , Michael Beyeler This is my paper
classification cs.LGcs.AI
keywords representation-structuredistillationbehaviortransferweak-to-stronggeneralizationout-of-distributionrobustnessCKAmodelalignmentknowledgeteacherselection
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 asks whether a model's aligned behavior—such as resistance to image corruptions—can be transferred to a different model trained on a different task, dataset, and architecture, without sharing the teacher's data. It claims yes, by distilling the geometry of the teacher's internal representations into the student: the student is fine-tuned to make its batch-level pairwise representation similarities match the teacher's, as measured by linear CKA, while still learning its own task. Across image classification transfers, BIRD outperforms fine-tuning, transfer learning, and continual learning baselines, improving robust accuracy by up to 16% over the strongest baseline. The paper also shows that a student up to $25\times$ larger than its teacher can be improved, and that three interpretable properties of the teacher's representations (task relevance, behavioral relevance, complementary knowledge) explain up to 85% of the variance in transfer success. If correct, BIRD turns small, well-aligned models into reusable alignment seeds for larger, heterogeneous students.

What carries the argument

The machinery is linear Centered Kernel Alignment (CKA), a similarity measure between two sets of representations computed on the same batch of inputs. For teacher representations $U$ and student representations $V$ on a batch $B$, the paper uses $\text{CKA}_{\text{linear}}(U,V) = \frac{\|V^T U\|_F^2}{\|U^T U\|_F \|V^T V\|_F}$ and defines the representation loss as $L_{\text{rep}} = 1 - \text{CKA}_{\text{linear}}(U,V)$. During fine-tuning, the student minimizes a weighted sum of its original task loss and this representation loss, with the teacher frozen. Because CKA compares pairwise similarity structure within batches rather than raw activations, the teacher and student need not share a label space, dataset, or architecture; the only inputs the teacher sees come from the student's distribution. The paper also builds two auxiliary metrics—task relevance (linear-probing accuracy and complementary knowledge) and behavioral relevance (aggregated $\gamma$-robust usefulness)—to predict which teacher layers transfer well.

What would settle it

If a teacher's feature vectors are randomly rotated (an orthogonal transformation, which leaves $\text{CKA}_{\text{linear}}$ unchanged) and BIRD's transferred robustness gains disappear, then pairwise similarity structure alone is not carrying the behavior, and the central claim fails. A concrete test: take a robust CIFAR-10 teacher, apply BIRD to a TinyImageNet student, and compare BIRD with and without a random orthogonal rotation applied to the teacher's features before distillation.

Watch

Extended reading notes

Core claim

The central claim is that aligned behavior can be transferred across models that differ in architecture, task, and training data by matching the internal representation structure of the student to that of the teacher. BIRD operationalizes this by minimizing $1 - \text{CKA}_{\text{linear}}$ between the pairwise similarity matrices of teacher and student representations computed on batches from the student's own input distribution, combined with the student's original task loss. The paper reports that this structure-matching objective transfers out-of-distribution robustness in image classification better than fine-tuning, linear probing, LP-FT, learning without forgetting, and activation-based hint distillation, even when the teacher is trained on a much simpler dataset and is $25\times$ smaller than the student. In a study of over 400 teacher--student pairs, three computable representation properties—task relevance, behavioral relevance, and complementary knowledge—explain up to 85% of the variance in transfer success. BIRD also gives complementary gains when added to soft-label weak-to-strong supervision on language tasks, though not in every configuration.

Load-bearing premise

The load-bearing premise is that the desired behavior is encoded in the geometry of a model's internal representations, and that a student forced to match that geometry on its own input distribution will inherit the behavior even though teacher and student were trained on different tasks and data.

Editorial extensions

If this is right

  • Small, well-aligned models can serve as reusable alignment seeds: a MobileNetV2 teacher trained on CIFAR-10 improves a ResNet-152 student on TinyImageNet, recovering 22.4% of the gap to a ceiling trained directly on corrupted data.
  • Because BIRD needs no teacher training data, aligned behavior can be transferred even when the teacher's dataset is private or proprietary, as long as the teacher accepts inputs from the student's domain.
  • Teacher selection becomes actionable: prioritizing teachers with high behavioral relevance—even with limited task overlap—should improve transfer, and training a small robust model on a curated dataset may be more efficient than robustly training a large one.
  • BIRD can complement soft-label distillation in language models: adding the representation-structure loss raised PGR from 17.7% to 24.2% when supervising GPT2-Large on SciQ, though gains are task-dependent.

Reading between the lines

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

  • Because linear CKA is invariant to orthogonal transformations, BIRD's success implies that the transferable behavioral signal lives in the pairwise similarity structure rather than in specific feature axes; a testable consequence is that randomly rotating the teacher's features should not change BIRD's outcome.
  • The same three predictors (task relevance, behavioral relevance, complementary knowledge) could be repurposed to select distillation layers within a single student, or to decide when a candidate teacher is too weak—the paper's saturation results for AlexNet teachers suggest a minimum teacher expressiveness exists for each target task.
  • If representation geometry is indeed the carrier of alignment, BIRD should also transfer other behaviors such as fairness or honesty with the same loss; a direct experiment would measure whether a teacher's representational signature of bias or dishonesty survives the same CKA distillation.
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 / 5 minor

Summary. The paper proposes BIRD, a method for transferring aligned behavior from a teacher model to a student model by matching the pairwise structure of their internal representations, measured by a linear CKA-like objective. The method is evaluated on out-of-distribution robustness transfer in image classification across three architectures and three dataset pairs, and on weak-to-strong soft-label distillation in NLP. The authors report that BIRD outperforms fine-tuning, linear probing, LP-FT, LwF, and activation-based Hints distillation, with robust-accuracy gains of up to 16% over the strongest baseline, and that three interpretable properties of teacher representations explain up to 85% of the variance in transfer success across 432 teacher-student pairs.

Significance. If the mechanism is sound, BIRD addresses a timely and important problem: reusing small, aligned models to induce behavioral properties in larger heterogeneous students without sharing teacher data or output spaces. The empirical scope is a strength, including 432 teacher-student pairs, held-out corruption types, per-seed tables for the vision experiments, and a large-scale teacher-selection analysis. The interpretable teacher-property analysis, if it were validated out of sample, would be practically valuable. However, the paper currently contains a formal error in the definition of the central objective, an in-sample predictive analysis, and high-variance NLP results that temper confidence in the stated claims.

major comments (4)
  1. [Section 3, Eq. (3)] The definition of CKA_linear in Eq. (3) is not the standard linear CKA. Standard linear CKA (Kornblith et al. [39]) has denominator ||u(B)^T u(B)||_F * ||v(B)^T v(B)||_F, not the product of the squared norms. As written, for u = v the value is 1/||u^T u||_F^2 rather than 1, so L_rep = 1 - CKA is not minimized by identical representation structure and is not scale-invariant. Minimizing this loss can be achieved by inflating the student representation norm, which does not align geometry. Because the paper's central claim is that BIRD transfers behavior by matching representation structure, this issue is load-bearing. Please correct Eq. (3) to the standard formula, or clarify the actual objective and implementation. If the implementation uses standard CKA, the reported results need to be verified under the stated objective; if it uses Eq. (3) as written, the improvements in Table 1 cannot be attributed to representation-structure matching.
  2. [Section 5, Figure 5] The R2 values in Figure 5 are computed in-sample: the linear model is fit to the same 432 teacher-student pairs whose PGR it is used to explain, and no cross-validation or held-out teacher set is reported. In-sample R2 with three predictors on 144 points per student dataset does not establish that transfer success is predictable for new teachers, which is the actionable claim of this section. Please report cross-validated R2 (e.g., leave-one-teacher-out or k-fold) and, ideally, a held-out set of teacher models.
  3. [Section 6, Tables 2 and 6] The averaged NLP results in Table 2 are not supported by the per-seed results in Table 6. For GPT2-Medium on CosmosQA, +BIRD is worse on average (42.76 vs. 47.00) and the per-seed values (7.00, 21.29, 100.00 for +BIRD vs. 16.94, 95.15, 28.89 for Soft-Label) show large variance and no consistent gain. For GPT2-Medium on SciQ, the +BIRD seeds are 0, 27.78, and 26.53 while Soft-Label seeds are 33.33, 0, and 12.24, so the reported average gain is driven by one seed. The claim that BIRD 'offers complementary gains' in Section 6 therefore needs to be tempered or supported by paired error bars and significance testing.
  4. [Appendix B.3] The baseline tuning protocol says that methods with additional hyperparameters are tuned over a range and the best performance is reported, but the text does not state whether alpha in Equation (1) was tuned for the vision experiments. If BIRD used a fixed default alpha while baselines were tuned to their best values, the comparison in Table 1 could be biased in favor of BIRD. Please report the alpha selection procedure for the vision experiments.
minor comments (5)
  1. [Appendix B.1] The list of corruption types used for teacher training contains 'brightness' twice; remove the duplicate.
  2. [Table 1] Table 1 reports means over three seeds without error bars; since Table 3 provides per-seed values, add standard deviations or confidence intervals to the main table.
  3. [Section 4, Figure 3] The text states that BIRD achieves higher robustness in every setting and corruption category, but the figure shows only three architectures and no numerical breakdown in the main text; specify the exact comparison and provide statistical support.
  4. [Section 5] The definition of R2 is not given in the main text; define it and state the number of points used in each fit.
  5. [Appendix F] The text contains the typo 'Cosmost QA' and should read 'Cosmos QA'.

Circularity Check

1 steps flagged · score 4.0 of 10

Core transfer benchmark is not circular, but the teacher-property 'explanation' is partly self-predictive because its key predictor is measured on the same target corruption benchmark as the outcome; Eq. 3 is a separate formal error, not a circularity.

  1. fitted input called prediction [Section 5, 'Behavioral relevance' and 'Explaining transfer success'; Appendix D.1]
    "Behavioral relevance We aggregate the γ-robust usefulness of each feature [51], which measures whether features retain predictive value under corruptions. ... For each student dataset, we fit a linear model to predict accuracy on the corrupted test set using the three metrics described above. ... We then evaluate the feature’s correlation with this binary labeling across target dataset images distorted with corruptions from the ImageNet-C benchmark."

    The 'behavioral relevance' predictor is computed on the target dataset under ImageNet-C corruptions, while the outcome it is used to explain is accuracy on the corrupted test set (PGR), which is evaluated on the same ImageNet-C corruption benchmark. The predictor is therefore a teacher-side version of the very corruption-robustness quantity the transfer is supposed to produce in the student, making the reported R^2 (73.6-85.5%) partly guaranteed by construction. The teacher-property analysis partially renames the teacher's own target-domain corruption robustness as an independent 'explanation' of transfer success, so it is a fitted-input-as-prediction pattern in the secondary analysis. The primary Table 1 benchmark remains external and non-circular.

full rationale

The central BIRD claim is not circular: robust accuracy is measured on held-out corruption types (speckle noise, gaussian blur, spatter, saturate) against five baseline methods, and the teacher and student differ in architecture, task, and data, so Table 1 results are externally anchored. No load-bearing self-citations appear; the cited NeuroAI and representation-learning work is external. The main circularity concern is Section 5's teacher-property analysis, where behavioral relevance (γ-robust usefulness under ImageNet-C corruptions on target-domain images) is used to 'predict' corrupted-test accuracy/PGR from the same corruption benchmark; the high R^2 is partly a construction artifact and should be presented as a teacher-side robustness measurement rather than an independent discovery. Separately, Eq. 3 defines a non-standard CKA: the denominator uses squared Frobenius norms, so for u=v the value is 1/||u^T u||_F^2 rather than 1, and L_rep is not minimized at identical structure; this is a formal correctness flaw in the stated objective, not a circular reduction, and it does not by itself make the benchmark circular. The α-selection protocol (Appendix F: 'selected based on performance') is under-specified; if it used the target metric on test data it would inflate results, but the text does not establish that, so it is not scored as a demonstrated circular step. Overall, the primary empirical comparison stands on its own, while the secondary explanatory claim is partly self-predictive, giving a moderate score of 4.

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

The central results rest on the hypothesis that behavior lives in representation geometry, on CKA as an adequate similarity measure across heterogeneous models, and on the validity of the three teacher-property proxies. The R2 analysis adds fitted regression coefficients and chosen aggregation thresholds. No new physical or conceptual entities are introduced.

free parameters (4)
  • alpha (representation loss weight) = not reported for vision; 5 values searched in NLP
    Eq. 1 controls the trade-off between task loss and CKA structure loss; Appendix F says it was selected based on performance from five pseudo-random values.
  • distillation anchor layer = Block 3 outputs in ResNets, similar transition points in other families
    Appendix B.4: layer chosen by a composite of linear-probing accuracy and gamma-robust usefulness, then fixed per model family.
  • gamma-robust usefulness aggregation thresholds = 90th percentile across classes, median across feature dimensions
    Appendix D.1: these hand-chosen thresholds shape the behavioral relevance metric used in the R2 analysis.
  • linear model coefficients for teacher-property predictors = not reported
    Section 5: a linear model is fit to predict PGR from three teacher metrics; the reported R2 values are in-sample and depend on these fitted coefficients.
assumptions (4)
  • domain assumption Behavioral properties such as robustness are encoded in the structure of internal representations, and matching that structure transfers the behavior.
    Section 3 opening: 'We posit that guiding a student model to adopt similar representational structure to its teacher biases it toward learning the same aligned behaviors.' This is the mechanism BIRD relies on.
  • domain assumption Linear CKA on batches from the student's input distribution is a valid and sufficient measure of structural similarity between heterogeneous teacher and student representations.
    Eq. 3 and Figure 1; assumes teacher forward passes on student-domain images produce meaningful geometry even with no shared training data.
  • domain assumption Linear probing accuracy and gamma-robust usefulness are valid proxies for task relevance and behavioral relevance of a teacher layer.
    Section 5 and Appendix D; these metrics are used as predictors in the R2 analysis and in the layer-selection heuristic.
  • domain assumption Robustness trained on 15 ImageNet-C corruption types transfers to the 4 held-out corruption types via representation structure.
    Section 4 and Appendix B.1; evaluation on held-out corruptions assumes the transferred signal is general robustness rather than memorized corruption appearance.

how reviews work

0 comments
Cite this review

Pith. "Pith review of BIRD: Behavior Induction via Representation-structure Distillation." pith.science (2026). https://pith.science/paper/VWFSNC7R

@misc{pith2026250523933,
  author       = {Pith},
  title        = {Pith review of: BIRD: Behavior Induction via Representation-structure Distillation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VWFSNC7R}},
  note         = {Machine review of arXiv:2505.23933}
}
abstract

Human-aligned deep learning models exhibit behaviors consistent with human values, such as robustness, fairness, and honesty. Transferring these behavioral properties to models trained on different tasks or data distributions remains challenging: aligned behavior is easily forgotten during fine-tuning, and collecting task-specific data that preserves this behavior can be prohibitively costly. We introduce BIRD (Behavior Induction via Representation-structure Distillation), a flexible framework for transferring aligned behavior by matching the internal representation structure of a student model to that of a teacher. Applied to out-of-distribution robustness in image classification, BIRD outperforms fine-tuning, transfer learning, and continual learning methods, improving robust accuracy by up to 16% over the next strongest baseline. It remains effective even when the teacher is trained on a much simpler dataset and is $25 \times$ smaller than the student. In a large-scale study of over 400 teacher-student pairs, we show that three interpretable and computable properties of the teacher's representations (i.e., task relevance, behavioral relevance, and complementary knowledge) explain up to 85% of the variance in transfer success. These insights offer practical guidance for teacher selection and design. BIRD turns small, well-aligned models into scalable alignment seeds, removing a key bottleneck in deploying safe AI systems in the wild.

Figures

Figures reproduced from arXiv: 2505.23933 by the authors.

Figure 1
Figure 1. Overview of BIRD: (a) First, a student model is pre-trained using generic optimization approaches on its target training set, Dstudent. A teacher model is independently trained to develop aligned behavior (e.g., robustness) on dataset Dteacher (optimizing an alignment objective, Lalign). (b) Next, the student model is fine-tuned on its original dataset to maintain task performance while learning latent representatio… view at source ↗
Figure 2
Figure 2. Sample image from CIFAR-10 with test-set corruptions from the ImageNet-C benchmark. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Comparing transferred robustness between BIRD and Hints students (CIFAR-10 trained [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (10 more)
Figure 4
Figure 4. Figure 4: OOD robustness of student models before and after BIRD, guided by a MobileNetV2 [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Variance in robust accuracy explained by linear models trained on teacher representation [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Sample CIFAR-10 image distorted with corruptions from the ImageNet-C benchmark [ [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]
Figure 7
Figure 7. Figure 7: Sample CIFAR-10 image distorted with “Speckle Noise” corruption at severities 1-5. [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]
Figure 8
Figure 8. Figure 8: Robust accuracy of CIFAR-100 student after applying BIRD with differing distillation [PITH_FULL_IMAGE:figures/full_fig_p016_8.png]
Figure 9
Figure 9. Figure 9: Performance gap recovered (PGR) of robust transfer methods applied to MobileNetV2, [PITH_FULL_IMAGE:figures/full_fig_p018_9.png]
Figure 10
Figure 10. Figure 10: Performance gap recovered (PGR) of robust transfer methods applied to ResNet18, by test [PITH_FULL_IMAGE:figures/full_fig_p019_10.png]
Figure 11
Figure 11. Figure 11: Performance gap recovered (PGR) of robust transfer methods applied to DenseNet169, [PITH_FULL_IMAGE:figures/full_fig_p019_11.png]
Figure 12
Figure 12. Figure 12: Change in per-class robust accuracy after fine-tuning with BIRD for each configuration of [PITH_FULL_IMAGE:figures/full_fig_p020_12.png]
Figure 13
Figure 13. Figure 13: Variance in robust accuracy explained by linear models trained on teacher representation [PITH_FULL_IMAGE:figures/full_fig_p020_13.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

59 extracted references · 21 canonical work pages

  1. [39]

    Similarity of neural network representations revisited

    Simon Kornblith, Mohammad Norouzi, Honglak Lee, and Geoffrey Hinton. Similarity of neural network representations revisited. InInternational conference on machine learning, pages 3519–3529. PMLR, 2019

    2019

  2. [1]

    Building safe artificial intelligence: specification, robustness, and assurance, 2018

    Pedro Ortega, Vishal Maini, and DeepMind Safety Team. Building safe artificial intelligence: specification, robustness, and assurance, 2018

    work page 2018

  3. [2]

    Ai alignment: A comprehensive survey.arXiv preprint arXiv:2310.19852, 2023

    Jiaming Ji, Tianyi Qiu, Boyuan Chen, Borong Zhang, Hantao Lou, Kaile Wang, Yawen Duan, Zhonghao He, Jiayi Zhou, Zhaowei Zhang, et al. Ai alignment: A comprehensive survey.arXiv preprint arXiv:2310.19852, 2023

    arXiv 2023

  4. [3]

    Towards deep learning models resistant to adversarial attacks.arXiv preprint arXiv:1706.06083, 2017

    Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks.arXiv preprint arXiv:1706.06083, 2017

    arXiv 2017

  5. [4]

    Training language models to follow instructions with human feedback.Advances in neural information processing systems, 35:27730–27744, 2022

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback.Advances in neural information processing systems, 35:27730–27744, 2022

    2022

  6. [5]

    Direct preference optimization: Your language model is secretly a reward model.Advances in Neural Information Processing Systems, 36:53728–53741, 2023

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model.Advances in Neural Information Processing Systems, 36:53728–53741, 2023

    2023

  7. [6]

    Adversarially robust transfer learning.arXiv preprint arXiv:1905.08232, 2019

    Ali Shafahi, Parsa Saadatpanah, Chen Zhu, Amin Ghiasi, Christoph Studer, David Jacobs, and Tom Goldstein. Adversarially robust transfer learning.arXiv preprint arXiv:1905.08232, 2019

    arXiv 1905

  8. [7]

    Fine- tuning aligned language models compromises safety, even when users do not intend to!arXiv preprint arXiv:2310.03693, 2023

    Xiangyu Qi, Yi Zeng, Tinghao Xie, Pin-Yu Chen, Ruoxi Jia, Prateek Mittal, and Peter Henderson. Fine- tuning aligned language models compromises safety, even when users do not intend to!arXiv preprint arXiv:2310.03693, 2023

    arXiv 2023

Show all 59 references
  1. [8]

    Navigating the safety landscape: Measuring risks in finetuning large language models.Advances in Neural Information Processing Systems, 37:95692– 95715, 2024

    Sheng Y Peng, Pin-Yu Chen, Matthew Hull, and Duen H Chau. Navigating the safety landscape: Measuring risks in finetuning large language models.Advances in Neural Information Processing Systems, 37:95692– 95715, 2024

    2024

  2. [9]

    Weak-to-strong generalization: Eliciting strong capabilities with weak supervision.arXiv preprint arXiv:2312.09390, 2023

    Collin Burns, Pavel Izmailov, Jan Hendrik Kirchner, Bowen Baker, Leo Gao, Leopold Aschenbrenner, Yining Chen, Adrien Ecoffet, Manas Joglekar, Jan Leike, et al. Weak-to-strong generalization: Eliciting strong capabilities with weak supervision.arXiv preprint arXiv:2312.09390, 2023

    2023 arXiv

  3. [10]

    Weak to strong generalization for large language models with multi-capabilities

    Yucheng Zhou, Jianbing Shen, and Yu Cheng. Weak to strong generalization for large language models with multi-capabilities. InThe Thirteenth International Conference on Learning Representations, 2025

    2025

  4. [11]

    On transfer of adversarial robustness from pretraining to downstream tasks.Advances in neural information processing systems, 36:59206– 59226, 2023

    Laura F Nern, Harsh Raj, Maurice André Georgi, and Yash Sharma. On transfer of adversarial robustness from pretraining to downstream tasks.Advances in neural information processing systems, 36:59206– 59226, 2023

    2023

  5. [12]

    When does contrastive learning preserve adversarial robustness from pretraining to finetuning?Advances in neural information processing systems, 34:21480–21492, 2021

    Lijie Fan, Sijia Liu, Pin-Yu Chen, Gaoyuan Zhang, and Chuang Gan. When does contrastive learning preserve adversarial robustness from pretraining to finetuning?Advances in neural information processing systems, 34:21480–21492, 2021

    2021

  6. [13]

    Neural population geometry: An approach for understanding biological and artificial neural networks.Current opinion in neurobiology, 70:137–144, 2021

    SueYeon Chung and Larry F Abbott. Neural population geometry: An approach for understanding biological and artificial neural networks.Current opinion in neurobiology, 70:137–144, 2021

    2021

  7. [14]

    Catalyzing next-generation artificial intelligence through neuroai.Nature communications, 14(1):1597, 2023

    Anthony Zador, Sean Escola, Blake Richards, Bence Ölveczky, Yoshua Bengio, Kwabena Boahen, Matthew Botvinick, Dmitri Chklovskii, Anne Churchland, Claudia Clopath, et al. Catalyzing next-generation artificial intelligence through neuroai.Nature communications, 14(1):1597, 2023

    2023

  8. [15]

    Neuroai for ai safety.arXiv preprint arXiv:2411.18526, 2024

    Patrick Mineault, Niccolò Zanichelli, Joanne Zichen Peng, Anton Arkhipov, Eli Bingham, Julian Jara- Ettinger, Emily Mackevicius, Adam Marblestone, Marcelo Mattar, Andrew Payne, et al. Neuroai for ai safety.arXiv preprint arXiv:2411.18526, 2024

    2024 arXiv

  9. [16]

    Joel Dapello, Kohitij Kar, Martin Schrimpf, Robert Baldwin Geary, Michael Ferguson, David Daniel Cox, and James J. DiCarlo. Aligning model and macaque inferior temporal cortex representations improves model-to-human behavioral alignment and adversarial robustness. InThe Eleven...

    2023

  10. [17]

    Towards robust vision by multi-task learning on monkey visual cortex.Advances in Neural Information Processing Systems, 34:739–751, 2021

    Shahd Safarani, Arne Nix, Konstantin Willeke, Santiago Cadena, Kelli Restivo, George Denfield, Andreas Tolias, and Fabian Sinz. Towards robust vision by multi-task learning on monkey visual cortex.Advances in Neural Information Processing Systems, 34:739–751, 2021

    2021

  11. [18]

    Learning from brains how to regularize machines

    Zhe Li, Wieland Brendel, Edgar Walker, Erick Cobos, Taliah Muhammad, Jacob Reimer, Matthias Bethge, Fabian Sinz, Zachary Pitkow, and Andreas Tolias. Learning from brains how to regularize machines. Advances in neural information processing systems, 32, 2019. 10

    2019

  12. [19]

    Representation engineering: A top-down approach to ai transparency.arXiv preprint arXiv:2310.01405, 2023

    Andy Zou, Long Phan, Sarah Chen, James Campbell, Phillip Guo, Richard Ren, Alexander Pan, Xuwang Yin, Mantas Mazeika, Ann-Kathrin Dombrowski, et al. Representation engineering: A top-down approach to ai transparency.arXiv preprint arXiv:2310.01405, 2023

    arXiv

  13. [20]

    Concrete problems in ai safety.arXiv preprint arXiv:1606.06565, 2016

    Dario Amodei, Chris Olah, Jacob Steinhardt, Paul Christiano, John Schulman, and Dan Mané. Concrete problems in ai safety.arXiv preprint arXiv:1606.06565, 2016

    2016 arXiv

  14. [21]

    Goal mis- generalization in deep reinforcement learning

    Lauro Langosco Di Langosco, Jack Koch, Lee D Sharkey, Jacob Pfau, and David Krueger. Goal mis- generalization in deep reinforcement learning. InInternational Conference on Machine Learning, pages 12004–12019. PMLR, 2022

    2022

  15. [22]

    Un- intentional unalignment: Likelihood displacement in direct preference optimization.arXiv preprint arXiv:2410.08847, 2024

    Noam Razin, Sadhika Malladi, Adithya Bhaskar, Danqi Chen, Sanjeev Arora, and Boris Hanin. Un- intentional unalignment: Likelihood displacement in direct preference optimization.arXiv preprint arXiv:2410.08847, 2024

    2024 arXiv

  16. [23]

    The many faces of robustness: A critical analysis of out-of-distribution generalization

    Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. InProceedings of the IEEE/CVF international co...

    2021

  17. [24]

    Weak-to-strong preference optimization: Stealing reward from weak aligned model.arXiv preprint arXiv:2410.18640, 2024

    Wenhong Zhu, Zhiwei He, Xiaofeng Wang, Pengfei Liu, and Rui Wang. Weak-to-strong preference optimization: Stealing reward from weak aligned model.arXiv preprint arXiv:2410.18640, 2024

    2024 arXiv

  18. [25]

    Cartl: Cooperative adversarially-robust transfer learning

    Dian Chen, Hongxin Hu, Qian Wang, Li Yinli, Cong Wang, Chao Shen, and Qi Li. Cartl: Cooperative adversarially-robust transfer learning. InInternational Conference on Machine Learning, pages 1640–1650. PMLR, 2021

    2021

  19. [26]

    Twins: A fine-tuning framework for improved transferability of adversarial robustness and generalization

    Ziquan Liu, Yi Xu, Xiangyang Ji, and Antoni B Chan. Twins: A fine-tuning framework for improved transferability of adversarial robustness and generalization. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16436–16446, 2023

    2023

  20. [27]

    Enhancing adversarial contrastive learning via adversarial invariant regularization.Advances in Neural Information Processing Systems, 36:16783–16803, 2023

    Xilie Xu, Jingfeng Zhang, Feng Liu, Masashi Sugiyama, and Mohan S Kankanhalli. Enhancing adversarial contrastive learning via adversarial invariant regularization.Advances in Neural Information Processing Systems, 36:16783–16803, 2023

    2023

  21. [28]

    Adversarially robust generalization requires more data.Advances in neural information processing systems, 31, 2018

    Ludwig Schmidt, Shibani Santurkar, Dimitris Tsipras, Kunal Talwar, and Aleksander Madry. Adversarially robust generalization requires more data.Advances in neural information processing systems, 31, 2018

    2018

  22. [29]

    Overcoming catastrophic forgetting in neural networks.Proceedings of the national academy of sciences, 114(13):3521–3526, 2017

    James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks.Proceedings of the national academy of scienc...

    2017

  23. [30]

    Learning without forgetting.IEEE transactions on pattern analysis and machine intelligence, 40(12):2935–2947, 2017

    Zhizhong Li and Derek Hoiem. Learning without forgetting.IEEE transactions on pattern analysis and machine intelligence, 40(12):2935–2947, 2017

    2017

  24. [31]

    Model compression

    Cristian Bucilua, Rich Caruana, and Alexandru Niculescu-Mizil. Model compression. InProceedings of the 12th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD ’06, page 535–541, New York, NY , USA, 2006. Association for Computing Machinery

    2006

  25. [32]

    Distilling the knowledge in a neural network.arXiv preprint arXiv:1503.02531, 2015

    Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network.arXiv preprint arXiv:1503.02531, 2015

    2015 arXiv

  26. [33]

    Fitnets: Hints for thin deep nets.arXiv preprint arXiv:1412.6550, 2014

    Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and Yoshua Bengio. Fitnets: Hints for thin deep nets.arXiv preprint arXiv:1412.6550, 2014

    2014 arXiv

  27. [34]

    Paying more attention to attention: Improving the performance of convolutional neural networks via attention transfer.arXiv preprint arXiv:1612.03928, 2016

    Sergey Zagoruyko and Nikos Komodakis. Paying more attention to attention: Improving the performance of convolutional neural networks via attention transfer.arXiv preprint arXiv:1612.03928, 2016

    2016 arXiv

  28. [35]

    Cross modal distillation for supervision transfer

    Saurabh Gupta, Judy Hoffman, and Jitendra Malik. Cross modal distillation for supervision transfer. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2827–2836, 2016

    2016

  29. [36]

    Learning deep representations by mutual information estimation and maximization

    R Devon Hjelm, Alex Fedorov, Samuel Lavoie-Marchildon, Karan Grewal, Phil Bachman, Adam Trischler, and Yoshua Bengio. Learning deep representations by mutual information estimation and maximization. arXiv preprint arXiv:1808.06670, 2018

    2018 arXiv

  30. [37]

    Contrastive representation distillation.arXiv preprint arXiv:1910.10699, 2019

    Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive representation distillation.arXiv preprint arXiv:1910.10699, 2019. 11

    1910 arXiv

  31. [38]

    Aligning machine and human visual representations across abstraction levels.arXiv preprint arXiv:2409.06509, 2024

    Lukas Muttenthaler, Klaus Greff, Frieda Born, Bernhard Spitzer, Simon Kornblith, Michael C Mozer, Klaus-Robert MÞller, Thomas Unterthiner, and Andrew K Lampinen. Aligning machine and human visual representations across abstraction levels.arXiv preprint arXiv:2409.06509, 2024

    2024 arXiv

  32. [40]

    Improved object recognition using neural networks trained to mimic the brain’s statistical properties.Neural Networks, 131:103–114, 2020

    Callie Federer, Haoyan Xu, Alona Fyshe, and Joel Zylberberg. Improved object recognition using neural networks trained to mimic the brain’s statistical properties.Neural Networks, 131:103–114, 2020

    2020

  33. [41]

    Learning multiple layers of features from tiny images

    Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009

    2009

  34. [42]

    Imagenet: A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In2009 IEEE conference on computer vision and pattern recognition, pages 248–255. Ieee, 2009

    2009

  35. [43]

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

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

    1903 arXiv

  36. [44]

    Mobilenetv2: Inverted residuals and linear bottlenecks

    Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhmoginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 4510–4520, 2018

    2018

  37. [45]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016

    2016

  38. [46]

    Densely connected convolutional networks

    Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 4700–4708, 2017

    2017

  39. [47]

    Fine-tuning can distort pretrained features and underperform out-of-distribution.arXiv preprint arXiv:2202.10054, 2022

    Ananya Kumar, Aditi Raghunathan, Robbie Jones, Tengyu Ma, and Percy Liang. Fine-tuning can distort pretrained features and underperform out-of-distribution.arXiv preprint arXiv:2202.10054, 2022

    2022 arXiv

  40. [48]

    Imagenet classification with deep convolutional neural networks

    Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In F. Pereira, C.J. Burges, L. Bottou, and K.Q. Weinberger, editors,Advances in Neural Information Processing Systems, volume 25. Curran Associates, Inc., 2012

    2012

  41. [49]

    Understanding intermediate layers using linear classifier probes

    Guillaume Alain and Yoshua Bengio. Understanding intermediate layers using linear classifier probes. arXiv preprint arXiv:1610.01644, 2016

    2016 arXiv

  42. [50]

    Fantastic gains and where to find them: On the existence and prospect of general knowledge transfer between any pretrained model.arXiv preprint arXiv:2310.17653, 2023

    Karsten Roth, Lukas Thede, Almut Sophia Koepke, Oriol Vinyals, Olivier Hénaff, and Zeynep Akata. Fantastic gains and where to find them: On the existence and prospect of general knowledge transfer between any pretrained model.arXiv preprint arXiv:2310.17653, 2023

    arXiv

  43. [51]

    Adversarial examples are not bugs, they are features.Advances in neural information processing systems, 32, 2019

    Andrew Ilyas, Shibani Santurkar, Dimitris Tsipras, Logan Engstrom, Brandon Tran, and Aleksander Madry. Adversarial examples are not bugs, they are features.Advances in neural information processing systems, 32, 2019

    2019

  44. [52]

    Huggingface’s transformers: State-of-the-art natural language processing.arXiv preprint arXiv:1910.03771, 2019

    Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, et al. Huggingface’s transformers: State-of-the-art natural language processing.arXiv preprint arXiv:1910.03771, 2019

    1910 arXiv

  45. [53]

    Matt Gardner Johannes Welbl, Nelson F. Liu. Crowdsourcing multiple choice science questions. 2017

    2017

  46. [54]

    Boolq: Exploring the surprising difficulty of natural yes/no questions

    Christopher Clark, Kenton Lee, Ming-Wei Chang, Tom Kwiatkowski, Michael Collins, and Kristina Toutanova. Boolq: Exploring the surprising difficulty of natural yes/no questions. InNAACL, 2019

    2019

  47. [55]

    Cosmos QA: Machine reading comprehension with contextual commonsense reasoning

    Lifu Huang, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. Cosmos QA: Machine reading comprehension with contextual commonsense reasoning. InProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference o...

    2019

  48. [56]

    gaussian noise

    Jesse Vig, Sebastian Gehrmann, Yonatan Belinkov, Sharon Qian, Daniel Nevo, Yaron Singer, and Stuart Shieber. Investigating gender bias in language models using causal mediation analysis.Advances in neural information processing systems, 33:12388–12401, 2020. 12 A Code availabi...

    2020

  49. [57]

    We then evaluate the feature’s correlation with this binary labeling across target dataset images distorted with corruptions from the ImageNet-C benchmark

    Compute class-wise γ-robust usefulness: For each feature dimension in the representation space, and for each class c in the target dataset, we binarize the class labels such that samples from class c are labeled as +1 and all other samples as −1. We then evaluate the feature’s...

  50. [58]

    This choice reflects the intuition that a useful feature may be strongly predictive of a subset of classes

    Feature-level aggregation: For each feature dimension, we compute the 90th percentile of its γ-robust usefulness scores across all classes. This choice reflects the intuition that a useful feature may be strongly predictive of a subset of classes

  51. [59]

    speckle-noise

    Layer-level aggregation: Finally, we aggregate over all feature dimensions by taking the median of the feature-level scores. This yields a single scalar score representing the typical γ-robust usefulness of features in the layer of interest. 16 Table 3: Per-seed accuracy (%) o...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.