Pith. sign in

REVIEW 3 major objections 6 minor 55 references

Vertical Fusion: Condensing Internal Representations for Robust ViT Classification

T0 review · 3 major / 6 minor · reviewed 2026-07-14 · grok-4.5

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

Pith's one-line read Intermediate Vision Transformer layers correct 18–76% of last-layer mistakes, and a learned vertical fusion of those layers closes 45% of the gap to an any-layer oracle.

desk verdict Solid multi-dataset recoverability numbers and a practical vertical fusion head; the diversity-vs-redundancy story is correlational, not causal, but the accuracy claims hold without it. read the letter →

arxiv 2607.10391 v1 pith:F5FSVYIO submitted 2026-07-11 cs.CV cs.LGeess.IV

Francesco Di Salvo , Shyam Nandan Rai , Hamed Damirchi , Ignacio Meza De la Jara , Sebastian Doerrich , Marco Lents , Christian Ledig This is my paper · ORCID
classification cs.CVcs.LGeess.IV
keywords VerticalEnsemblingInternalRepresentationsRecoverabilityVisionTransformersFeatureFusionLayerAggregationOut-of-DistributionClassification
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

Vision Transformers are almost always used as black boxes that expose only their final layer for classification. This paper shows that is wasteful: when a lightweight probe is trained on every intermediate layer across 16 datasets, those probes recover between 18% and 76% of the samples the final layer gets wrong. The authors argue the gains come less from predictive diversity than from a redundancy-correctness pattern—layers act as slightly different probes of a shared discriminative signal. They therefore introduce VFusion, a small learned encoder that concatenates selected layer tokens, compresses them into a low-dimensional latent, and trains a classifier on that latent while the backbone stays frozen. The result beats standard aggregation baselines on both in-distribution and shifted data, and works across model sizes and pre-training styles, offering a cheap vertical alternative to multi-model ensembles.

What carries the argument

VFusion: a learnable fusion encoder that concatenates selected intermediate [CLS] tokens, maps the concatenation into a low-dimensional latent (with optional orthogonalization), and trains a lightweight classifier on that latent while the backbone remains frozen; together with the recoverability rate (fraction of final-layer errors corrected by at least one intermediate probe).

What would settle it

On a held-out collection of classification datasets, measure layer-wise recovery rates and train VFusion against best-layer and aggregation baselines; if recovery rates collapse toward zero and VFusion no longer beats the best single layer or the oracle gap remains largely unclosed, the central claim fails.

Watch

Extended reading notes

Core claim

Intermediate representations inside a single frozen Vision Transformer contain substantial corrective signal: independent layer-wise probes recover 18–76% of last-layer errors across 16 datasets. That recoverability is better explained by distributed, redundant probes of a shared decision boundary than by ensemble-style disagreement. A lightweight supervised fusion head (VFusion) that compresses the multi-layer hierarchy into one low-dimensional token captures a large fraction of this unused signal, closing 45% of the accuracy gap between the best single layer and a theoretical any-layer oracle, and outperforming established aggregation methods in both in-distribution and out-of-distribution

Load-bearing premise

That the corrective power of intermediate layers comes mainly from them being stable, redundant probes of one shared signal rather than from genuine predictive diversity, a claim resting on correlations with correction entropy and logit disagreement rather than a direct causal test.

Editorial extensions

If this is right

  • A frozen off-the-shelf ViT can be made more accurate by attaching only a small fusion head instead of training or running multiple backbones.
  • Horizontal ensembles become less necessary when the vertical hierarchy inside one model already supplies recoverable signal.
  • Deeper contiguous layers preserve most of the fusion benefit, so practitioners can safely drop early layers if compute is tight.
  • The same compression principle extends to horizontal fusion across heterogeneous pre-trained backbones when multiple models are available.
  • Gains appear larger on fine-grained and degraded-image tasks, suggesting the method is most useful where last-layer features are fragile.

Reading between the lines

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

  • The same recoverability pattern may appear in other transformer stacks (language, multimodal) that keep relatively uniform representations across depth, inviting vertical fusion heads beyond vision classification.
  • If the redundancy-correctness account holds, unsupervised or self-supervised pre-training of the fusion encoder could remove the need for task labels at fusion time.
  • Dense prediction tasks such as segmentation may benefit even more, because intermediate layers already encode spatial detail that the final [CLS] token discards.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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. The paper argues that intermediate Vision Transformer layers contain substantial corrective signal for multi-class classification: independent layer-wise probes recover 18%–76% of last-layer errors across 16 datasets (Table 1; recovery rate in Eq. 2). It attributes this primarily to a “redundancy-correctness correspondence” rather than predictive diversity, based on correlations of recovery rate with correction entropy and (negatively) with pairwise logit disagreement (Fig. 1; §3.2). Building on that analysis, it proposes VFusion: select multi-layer [CLS] tokens, concatenate them, and compress them with a supervised MLP encoder into a low-dimensional latent (Eqs. 6–7) trained with cross-entropy plus an orthogonalization regularizer (Eqs. 8–10). Empirically, VFusion outperforms non-parametric and parametric aggregation baselines (Average, Majority, Super Learner, NLC, MoE, sw-MoE) on 16 ID datasets (Table 2; avg. 91.4% vs best-layer 89.5%, closing ~45% of the best-layer–oracle gap) and five shift benchmarks (Table 3), with ablations on layer selection, backbone family/scale, horizontal fusion, capacity, efficiency, and latent size.

Significance. If the empirical results hold, the work is a useful and practical contribution to frozen-backbone ViT deployment: it quantifies internal recoverability at scale, shows that simple last-layer probing leaves unused signal, and provides a lightweight vertical fusion head that improves ID and OOD accuracy without multi-backbone ensembles. Strengths include broad public-benchmark evaluation (21 datasets), multiple established baselines, capacity-controlled Final-only comparison (Appendix C.2), efficiency vs horizontal Super Learner (Appendix C.1), generalization across supervised/self-supervised/CLIP backbones and scales (Fig. 3), and released code. The recoverability framing and layer-selection guideline (contiguous depth preferred over striding) are actionable for practitioners. The mechanistic “redundancy-correctness” story is softer than the accuracy claims and should be treated as interpretive rather than established; even so, the method and measurement contributions remain significant for the field.

major comments (3)
  1. [§3.2, Fig. 1, Abstract] §3.2 and Fig. 1 (also Abstract/Introduction): the claim that recoverability is “not primarily driven by predictive diversity, but by a redundancy-correctness correspondence” is supported only by observational correlations (recovery rate vs H_corr, r=0.87; vs average pairwise cosine distance on jointly misclassified samples, r=−0.77; JS check in Appendix A.1). These do not causally isolate shared-signal compression from residual local complementarity (Appendix A.2 documents non-zero unique neighbor recoveries at every depth). Please rephrase causal language to correlational evidence, and clarify that VFusion’s compression design is motivated by—not proven by—this analysis. The accuracy claims do not require a stronger mechanism proof, but the current wording overstates what Fig. 1 establishes.
  2. [§4.2.1, Eqs. (6)–(7)] §4.2.1 / Eqs. (6)–(7): the decomposition h^(ℓ)=s^(ℓ)+r^(ℓ) and the assertion that the low-dimensional encoder “discards layer-specific residuals” are motivational, not measured. No ablation isolates shared vs layer-private components (e.g., reconstruction of s, mutual information across layers, or comparison of low-dim fusion vs a capacity-matched high-dim concat MLP without aggressive bottleneck). Appendix C.2 shows multi-layer input helps vs Final-only, and MoE uses the same concat features, but neither confirms noise-suppression of r^(ℓ). Either add a targeted diagnostic or present the encoder as supervised multi-layer dimensionality reduction without claiming residual filtering as established.
  3. [Table 2, §5.2, Eq. (1)] Table 2 / §5.2: VFusion approaches or exceeds the layer-wise oracle on Cars, Flow, IN1k, and ESAT. The oracle (Eq. 1) marks a sample correct if any single-layer probe is correct; surpassing it implies the fused representation creates new decision boundaries, not only selection among layers. This is interesting but under-discussed. Please quantify how often VFusion is correct when all layer probes fail (or when only a minority succeed), and state clearly that the oracle is not an upper bound on feature-level fusion—only on discrete layer selection—so the “45% of headroom closed” framing does not imply proximity to an absolute ceiling.
minor comments (6)
  1. [Tables 1–2] Tables consistently typeset “ESA T” / “ESA T” with a space (EuroSAT). Fix throughout Tables 1–2 and related captions.
  2. [§5.1.3] §5.1.3: probe hyperparameter grid and VFusion settings are clear; please also state whether MoE/sw-MoE expert MLPs match the probe head capacity and whether routing is trained jointly with experts under the same early-stopping protocol.
  3. [Figure 2] Figure 2 right panel is effective; ensure the “45%” annotation is defined in the caption as (a_VFusion − a_best)/(a_oracle − a_best) using the Table 2 averages so the figure is self-contained.
  4. [§6.3, Table 5] §6.3 / Table 5: HFusion is introduced as horizontal feature fusion; a one-sentence architectural parallel to VFusion (same encoder family on concatenated last-layer features) would help readers map vertical vs horizontal settings.
  5. [Appendix A.2] Appendix A.2 Table A1 is informative; consider a brief pointer in main §3.2 that local unique recoveries exist at all depths, so redundancy is imperfect—this would balance the low-disagreement narrative without new experiments.
  6. [§7.1] Limitations (§7.1) correctly note supervised fusion and frozen backbones; a short remark that recoverability is measured with trained probes (not zero-shot) would prevent over-reading for CLIP-style settings already discussed in §2.3.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: recoverability and VFusion gains are measured against external benchmarks and independent baselines, not forced by definition or self-citation.

full rationale

This is an empirical methods paper. Recoverability (Eq. 2) is defined as the fraction of last-layer errors corrected by any intermediate probe and is measured on held-out test sets across 16 public datasets (Table 1); the oracle (Eq. 1) is an upper bound, not a fitted target that VFusion is forced to match. VFusion is a learned encoder+head trained with cross-entropy plus an optional orthogonalization regularizer (Eqs. 7–10) on concatenated frozen [CLS] tokens; its reported accuracy (Tables 2–3) is compared to independent baselines (Best layer, Average, Majority, SL, NLC, MoE, sw-MoE) and a capacity-controlled Final-only ablation (Appendix C.2). Closing 45% of the best-layer-to-oracle gap is an empirical observation, not an algebraic identity. The redundancy-correctness story (Fig. 1 correlations of recovery rate with correction entropy and negative correlation with logit cosine/JS disagreement) is interpretive support for the design choice of compression over diversity; it does not define the accuracy metric or force the numbers. No uniqueness theorem, ansatz smuggled via self-citation, or fitted parameter renamed as prediction appears. Self-citations are ordinary related-work references and are not load-bearing for the central claims. Score 0 is appropriate.

Assumptions & free parameters 4 free parameters · 4 assumptions · 3 invented entities

The work is empirical: claims rest on standard frozen-backbone transfer assumptions, CLS-token probes, and a handful of hand-chosen fusion hyperparameters, plus the interpretive axiom that low disagreement plus high correction entropy means redundancy-correctness. No new physical entities; invented constructs are methodological (recoverability, VFusion, the redundancy-correctness framing).

free parameters (4)
  • latent dimension dz
    Default dz=256 chosen for the fusion bottleneck; appendix sweep shows limited sensitivity but the main results use this fixed capacity.
  • orthogonalization weight λ
    Default λ=2.0 in L_total = L_CE + λ L_orth; swept in appendix but not derived from theory.
  • fusion encoder widths [2048,1024,512]
    Architecture of E is a design choice that sets trainable capacity of the vertical head.
  • probe and fusion optim hyperparameters
    Learning rates, weight decay, batch size 512, early-stopping patience, and per-layer probe grid search affect reported recoverability and accuracy.
assumptions (4)
  • domain assumption Layer-wise [CLS] tokens are adequate global descriptors of intermediate ViT representations for classification probes and fusion.
    §3.1 defines h^(ℓ) as the layer-ℓ [CLS] token and builds all recoverability and VFusion analysis on that choice.
  • domain assumption A frozen pretrained backbone plus a supervised lightweight head is a valid setting for measuring and exploiting internal corrective signal.
    Throughout §4–5 the backbone Φ is frozen; limitations note fine-tuning interactions are unexplored.
  • ad hoc to paper h^(ℓ) = s^(ℓ) + r^(ℓ) with a low-dimensional encoder isolating shared task-relevant s across layers.
    §4.2.1–4.2.2 introduces this decomposition as motivation for compression; it is a modeling story, not independently measured.
  • ad hoc to paper Recovery rate and correction entropy quantify useful internal corrective capacity for multi-class classification.
    Definitions (1)–(4) in §3.2 operationalize recoverability; the paper’s narrative treats high recovery as evidence of actionable vertical signal.
invented entities (3)
  • recoverability / recovery rate independent evidence
    purpose: Quantify how often intermediate probes correct last-layer errors, normalized by last-layer error mass.
    Defined in §3.2.1; central diagnostic motivating vertical fusion.
  • redundancy-correctness correspondence
    purpose: Interpret recoverability as stable redundant probes of a shared signal rather than ensemble diversity.
    Inferred from Fig. 1 correlations; not a separately validated physical object.
  • VFusion fusion encoder + orthogonalized latent global token independent evidence
    purpose: Compress multi-layer features into a task-specific latent for classification.
    Proposed method in §4; evaluated empirically against baselines.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Vertical Fusion: Condensing Internal Representations for Robust ViT Classification." pith.science (2026). https://pith.science/paper/F5FSVYIO

@misc{pith2026260710391,
  author       = {Pith},
  title        = {Pith review of: Vertical Fusion: Condensing Internal Representations for Robust ViT Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/F5FSVYIO}},
  note         = {Machine review of arXiv:2607.10391}
}
read the original abstract

Despite exposing rich intermediate representations, Vision Transformers (ViTs) are almost exclusively utilized as black-box feature extractors, where only the last layer is considered for downstream tasks. We challenge this convention by introducing the notion of recoverability: the capacity of intermediate representations to correct last-layer failures. By evaluating independent classification probes at every model depth across 16 datasets, we observe that intermediate probes correctly classify 18% to 76% of samples that the last-layer probe misclassifies. We show that these gains are not primarily driven by predictive diversity, but by a redundancy-correctness correspondence, where the internal hierarchy acts as a series of stable, redundant probes of a shared discriminative signal. While established horizontal ensemble strategies (i.e., across multiple models) can improve performance, they incur high computational cost and ignore this vertical signal within a single model. To bridge this gap, we propose VFusion, a principled vertical aggregation strategy employing a learnable mapping into a low-dimensional latent space that synthesizes features across the internal ViT hierarchy. VFusion substantially outperforms established aggregation baselines in both in-distribution and out-of-distribution settings, notably closing 45% of the accuracy gap between the best individual layer and a theoretical oracle performance. Our gains consistently generalize across model sizes and pre-training regimes, confirming that VFusion offers a robust and efficient alternative to horizontal ensemble methods. The code is available at https://github.com/francescodisalvo05/vit-vertical-fusion.

Figures

Figures reproduced from arXiv: 2607.10391 by the authors.

Figure 1
Figure 1. We measure correction entropy Hcorr, defined as the entropy of the conditional correction prob￾abilities cℓ = P(ℓ correct | L wrong), which quantifies how distributed correction capacity is across inter￾mediate layers ℓ. Higher entropy indicates that multiple layers contribute complementary boundary-level corrections. The recovery rate (%), i.e., fraction of last-layer (L) errors recoverable by intermediate layers, … view at source ↗
Figure 2
Figure 2. Left: Illustration of VFusion. A frozen ViT backbone Φ extracts layer representations {h (ℓ) } L ℓ=1 , with h (ℓ) = s (ℓ) + r (ℓ) (task-relevant + task-irrelevant). A subset of M layers is selected and concatenated into v = [h (ℓ) ]ℓ∈S ∈ R M·d , which is compressed by a learnable fusion encoder E into a low-dimensional latent z ∈ R dz (dz ≪ M · d). A lightweight head Γ predicts class logits from z using cross-entrop… view at source ↗
Figure 3
Figure 3. Average improvement over best-layer performance ( [PITH_FULL_IMAGE:figures/full_fig_p019_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

55 extracted references · 3 linked inside Pith

  1. [1]

    Dosovitskiy, L

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Un- terthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, N. Houlsby, An image is worth 16x16 words: Transformers for image recog- nition at scale, in: International Conference on Learning Representations, 2021

    2021

  2. [2]

    Ranftl, A

    R. Ranftl, A. Bochkovskiy, V . Koltun, Vision transformers for dense prediction, in: Proceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 12179–12188. 23

    2021

  3. [3]

    Oquab, T

    M. Oquab, T. Darcet, T. Moutakanni, H. V . V o, M. Szafraniec, V . Khalidov, P. Fernandez, D. HAZIZA, F. Massa, A. El-Nouby, M. Assran, N. Ballas, W. Galuba, R. Howes, P.-Y . Huang, S.-W. Li, I. Misra, M. Rabbat, V . Sharma, G. Synnaeve, H. Xu, H. Jegou, J. Mairal, P. Labatut, A. Joulin, P. Bojanowski, DINOv2: Learning robust visual features without super...

    2024

  4. [4]

    Teterwak, K

    P. Teterwak, K. Saito, T. Tsiligkaridis, B. A. Plummer, K. Saenko, Is large- scale pretraining the secret to good domain generalization?, in: The Thirteenth International Conference on Learning Representations, 2025

    2025

  5. [5]

    R. Imam, R. Marew, M. Yaqub, On the robustness of medical vision-language models: Are they truly generalizable?, in: Annual Conference on Medical Image Understanding and Analysis, Springer, 2025, pp. 233–256

    2025

  6. [6]

    Kumar, N

    D. Kumar, N. Muhammad, Object detection in adverse weather for autonomous driving through data merging and yolov8, Sensors 23 (20) (2023) 8471

    2023

  7. [7]

    W. He, Z. Jiang, T. Xiao, Z. Xu, Y . Li, A survey on uncertainty quantification methods for deep learning, ACM Computing Surveys 58 (7) (2025)

    2025

  8. [8]

    Minderer, J

    M. Minderer, J. Djolonga, R. Romijnders, F. Hubis, X. Zhai, N. Houlsby, D. Tran, M. Lucic, Revisiting the calibration of modern neural networks, Ad- vances in neural information processing systems 34 (2021) 15682–15694

    2021

Show all 55 references
  1. [9]

    Pinto, P

    F. Pinto, P. H. Torr, P. K. Dokania, An impartial take to the cnn vs transformer robustness contest, in: European conference on computer vision, Springer, 2022

    2022

  2. [10]

    Gulrajani, D

    I. Gulrajani, D. Lopez-Paz, In search of lost domain generalization, in: Interna- tional Conference on Learning Representations, 2021

    2021

  3. [11]

    Hendrycks, T

    D. Hendrycks, T. Dietterich, Benchmarking neural network robustness to com- mon corruptions and perturbations, Proceedings of the International Conference on Learning Representations (2019). 24

    2019

  4. [12]

    Raghu, T

    M. Raghu, T. Unterthiner, S. Kornblith, C. Zhang, A. Dosovitskiy, Do vision transformers see like convolutional neural networks?, Advances in neural infor- mation processing systems 34 (2021) 12116–12128

    2021

  5. [13]

    Wei, B.-L

    T. Wei, B.-L. Wang, J.-X. Shi, Y .-F. Li, M.-L. Zhang, X-mahalanobis: Trans- former feature mixing for reliable OOD detection, in: The Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025

    2025

  6. [14]

    Jeleni ´c, J

    F. Jeleni ´c, J. Juki ´c, M. Tutek, M. Puljiz, J. Snajder, Out-of-distribution detec- tion by leveraging between-layer transformation smoothness, in: The Twelfth International Conference on Learning Representations, 2024

    2024

  7. [15]

    Rodriguez-Opazo, D

    Imezadelajara, C. Rodriguez-Opazo, D. Teney, D. Ranasinghe, E. Abbasnejad, Mysteries of the deep: Role of intermediate representations in out of distribu- tion detection, in: The Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025

    2025

  8. [16]

    Uselis, S

    A. Uselis, S. J. Oh, Intermediate layer classifiers for OOD generalization, in: The Thirteenth International Conference on Learning Representations, 2025

    2025

  9. [17]

    Rodriguez-Opazo, E

    C. Rodriguez-Opazo, E. Abbasnejad, D. Teney, H. Damirchi, E. Marrese-Taylor, A. van den Hengel, Synergy and diversity in CLIP: Enhancing performance through adaptive backbone ensembling, in: The Thirteenth International Con- ference on Learning Representations, 2025

    2025

  10. [18]

    Radford, J

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, et al., Learning transferable visual models from natural language supervision, in: International Conference on Machine Learning (ICML), PmLR, 2021, pp. 8748–8763

    2021

  11. [19]

    Nakata, Y

    K. Nakata, Y . Ng, D. Miyashita, A. Maki, Y .-C. Lin, J. Deguchi, Revisiting a knn-based image classification system with high-capacity storage, in: European conference on computer vision, Springer, 2022, pp. 457–474

    2022

  12. [20]

    Di Salvo, S

    F. Di Salvo, S. Doerrich, I. Rieger, C. Ledig, An embedding is worth a thousand noisy labels, Transactions on Machine Learning Research (2025). 25

    2025

  13. [21]

    Y . Zhu, J. Zhang, A. Gangrade, C. Scott, Label noise: Ignorance is bliss, Ad- vances in Neural Information Processing Systems 37 (2024) 116575–116616

    2024

  14. [22]

    Mayilvahanan, R

    P. Mayilvahanan, R. S. Zimmermann, T. Wiedemer, E. Rusak, A. Juhos, M. Bethge, W. Brendel, In search of forgotten domain generalization, in: The Thirteenth International Conference on Learning Representations, 2025

    2025

  15. [23]

    Hendrycks, K

    D. Hendrycks, K. Gimpel, A baseline for detecting misclassified and out-of- distribution examples in neural networks, in: International Conference on Learn- ing Representations, 2017

    2017

  16. [24]

    Y . Sun, Y . Ming, X. Zhu, Y . Li, Out-of-distribution detection with deep nearest neighbors, International Conference on Machine Learning (ICML) (2022)

    2022

  17. [25]

    Koutlis, S

    C. Koutlis, S. Papadopoulos, Leveraging representations from intermediate encoder-blocks for synthetic image detection, in: European Conference on Com- puter Vision, Springer, 2024, pp. 394–411

    2024

  18. [26]

    T. G. Dietterich, Ensemble methods in machine learning, in: International work- shop on multiple classifier systems (MCS), Springer, 2000, pp. 1–15

    2000

  19. [27]

    Freund, R

    Y . Freund, R. E. Schapire, A decision-theoretic generalization of on-line learning and an application to boosting, Journal of computer and system sciences 55 (1) (1997) 119–139

    1997

  20. [28]

    Lakshminarayanan, A

    B. Lakshminarayanan, A. Pritzel, C. Blundell, Simple and scalable predictive uncertainty estimation using deep ensembles, Advances in neural information processing systems 30 (2017)

    2017

  21. [29]

    S. Fort, H. Hu, B. Lakshminarayanan, Deep ensembles: A loss landscape per- spective, arXiv preprint arXiv:1912.02757 (2019)

    1912 arXiv

  22. [30]

    Wortsman, G

    M. Wortsman, G. Ilharco, S. Y . Gadre, R. Roelofs, R. Gontijo-Lopes, A. S. Morcos, H. Namkoong, A. Farhadi, Y . Carmon, S. Kornblith, et al., Model soups: averaging weights of multiple fine-tuned models improves accuracy with- out increasing inference time, in: International c...

    2022

  23. [31]

    Ainsworth, J

    S. Ainsworth, J. Hayase, S. Srinivasa, Git re-basin: Merging models modulo permutation symmetries, in: The Eleventh International Conference on Learning Representations, 2023

    2023

  24. [32]

    K. Lenc, A. Vedaldi, Understanding image representations by measuring their equivariance and equivalence, in: Proceedings of the IEEE conference on com- puter vision and pattern recognition, 2015, pp. 991–999

    2015

  25. [33]

    Bansal, P

    Y . Bansal, P. Nakkiran, B. Barak, Revisiting model stitching to compare neural representations, Advances in neural information processing systems 34 (2021)

    2021

  26. [34]

    C. Ju, A. Bibaut, M. van der Laan, The relative performance of ensemble meth- ods with deep convolutional neural networks for image classification, Journal of applied statistics 45 (15) (2018) 2800–2818

    2018

  27. [35]

    Zbontar, L

    J. Zbontar, L. Jing, I. Misra, Y . LeCun, S. Deny, Barlow twins: Self-supervised learning via redundancy reduction, in: International conference on machine learning, PMLR, 2021, pp. 12310–12320

    2021

  28. [36]

    Bardes, J

    A. Bardes, J. Ponce, Y . LeCun, VICReg: Variance-invariance-covariance regu- larization for self-supervised learning, in: International Conference on Learning Representations, 2022

    2022

  29. [37]

    Krizhevsky, G

    A. Krizhevsky, G. Hinton, et al., Learning multiple layers of features from tiny images (2009)

    2009

  30. [38]

    Fei-Fei, R

    L. Fei-Fei, R. Fergus, P. Perona, Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object cate- gories, Computer vision and Image understanding 106 (1) (2007) 59–70

    2007

  31. [39]

    Krause, M

    J. Krause, M. Stark, J. Deng, L. Fei-Fei, 3d object representations for fine- grained categorization, in: Proceedings of the IEEE international conference on computer vision workshops, 2013, pp. 554–561

    2013

  32. [40]

    C. Wah, S. Branson, P. Welinder, P. Perona, S. Belongie, The caltech-ucsd birds- 200-2011 dataset (2011). 27

    2011

  33. [41]

    Cimpoi, S

    M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, A. Vedaldi, Describing textures in the wild, in: Proceedings of the IEEE conference on computer vision and pattern recognition, 2014, pp. 3606–3613

    2014

  34. [42]

    Helber, B

    P. Helber, B. Bischke, A. Dengel, D. Borth, Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification, IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 12 (7) (2019) 2217–2226

    2019

  35. [43]

    S. Maji, E. Rahtu, J. Kannala, M. Blaschko, A. Vedaldi, Fine-grained visual classification of aircraft, arXiv preprint arXiv:1306.5151 (2013)

    2013 arXiv

  36. [44]

    Nilsback, A

    M.-E. Nilsback, A. Zisserman, Automated flower classification over a large num- ber of classes, in: Indian Conference on Computer Vision, Graphics and Image Processing, 2008

    2008

  37. [45]

    Bossard, M

    L. Bossard, M. Guillaumin, L. Van Gool, Food-101 – mining discriminative components with random forests, in: European Conference on Computer Vision, 2014

    2014

  38. [46]

    Houben, J

    S. Houben, J. Stallkamp, J. Salmen, M. Schlipsing, C. Igel, Detection of traffic signs in real-world images: The German Traffic Sign Detection Benchmark, in: International Joint Conference on Neural Networks, no. 1288, 2013

    2013

  39. [47]

    LeCun, C

    Y . LeCun, C. Cortes, C. Burges, Mnist handwritten digit database, ATT Labs [Online]. Available: http://yann.lecun.com/exdb/mnist 2 (2010)

    2010

  40. [48]

    B. S. Veeling, J. Linmans, J. Winkens, T. Cohen, M. Welling, Rotation equiv- ariant cnns for digital pathology, in: International Conference on Medical image computing and computer-assisted intervention, Springer, 2018, pp. 210–218

    2018

  41. [49]

    O. M. Parkhi, A. Vedaldi, A. Zisserman, C. Jawahar, Cats and dogs, in: 2012 IEEE conference on computer vision and pattern recognition, IEEE, 2012, pp. 3498–3505. 28

    2012

  42. [50]

    Coates, A

    A. Coates, A. Ng, H. Lee, An analysis of single-layer networks in unsupervised feature learning, in: Proceedings of the fourteenth international conference on ar- tificial intelligence and statistics, Journal of Machine Learning Research (JMLR) Workshop and Conference Proceedi...

    2011

  43. [51]

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, L. Fei-Fei, Imagenet: A large-scale hierarchical image database, in: 2009 IEEE Conference on Computer Vision and Pattern Recognition, 2009, pp. 248–255.doi:10.1109/CVPR.2009.5206848

    2009 doi

  44. [52]

    Z. Liu, P. Luo, X. Wang, X. Tang, Deep learning face attributes in the wild, in: Proceedings of the IEEE international conference on computer vision, 2015, pp. 3730–3738

    2015

  45. [53]

    Sagawa, P

    S. Sagawa, P. W. Koh, T. B. Hashimoto, P. Liang, Distributionally robust neural networks, in: International Conference on Learning Representations, 2020

    2020

  46. [54]

    Shazeer, A

    N. Shazeer, A. Mirhoseini, K. Maziarz, A. Davis, Q. Le, G. Hinton, J. Dean, Out- rageously large neural networks: The sparsely-gated mixture-of-experts layer, arXiv preprint arXiv:1701.06538 (2017)

    2017 arXiv

  47. [55]

    Fedus, B

    W. Fedus, B. Zoph, N. Shazeer, Switch transformers: Scaling to trillion pa- rameter models with simple and efficient sparsity, Journal of Machine Learning Research 23 (120) (2022) 1–39. 29 Vertical Fusion: Condensing Internal Representations for Robust ViT Classification Appen...

    2022

Pith tools

Reviewed July 14, 2026 · model on record in the stance chip above.