Pith. sign in

REVIEW 4 major objections 5 minor 35 references

The paper proves feedforward networks are a strict subset of convolutional networks and introduces model projection, a method that lets CNNs train like FFNs with one weight per input channel and transfer efficiently to downstream tasks.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 03:58 UTC pith:OBAHFRNB

load-bearing objection A clear, honest paper whose main theorem restates the 1×1-convolution equivalence and whose practical claim is not yet supported by matched comparisons; the model-projection method itself is worth a serious look. the 4 major comments →

arxiv 2602.06245 v2 pith:OBAHFRNB submitted 2026-02-05 stat.ML cs.LG

Inheritance Between Feedforward and Convolutional Networks via Model Projection

Nicolas Ewen , Jairo Diaz-Rodriguez , Kelly Ramsay This is my paper
classification stat.ML cs.LG MSC 68T07
keywords model projectioninheritance between model classesgeneralized feedforward networksgeneralized convolutional networksparameter-efficient transfer learningseparable by inputtensor dot product
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper asks when techniques developed for one neural-network family can legitimately be transferred to another. It answers with a unified node-level formalization in which generalized feedforward networks (GFFNs) are a strict subset of generalized convolutional networks (GCNNs), so results that hold for all GCNNs automatically hold for GFFNs. The reverse direction fails in general, but the paper introduces model projection: freeze each pretrained convolutional filter and learn one scalar per input-output channel pair, which turns a CNN node into a GFFN-style node with inhomogeneous inputs. This lets projected CNNs inherit FFN training techniques, such as one weight per input channel, and yields a parameter-efficient transfer learning method that is competitive with established PEFT baselines and serves as an effective initialization for full fine-tuning.

Core claim

With tensor-valued activations, a GFFN node computes σ(Σ Z_k w_k + b) and a GCNN node σ(Σ Z_k * F_k + b). A kernel of size one makes convolution into scalar multiplication, so GFFNs are exactly the GCNNs with 1×1 kernels — a strict subclass (Theorem 3.5). Model projection freezes per-channel sub-functions and learns one scalar γ per output/input channel pair, giving σ(Σ γ_{jk} f'_{jk}(Z_k) + b). Theorem 3.11 proves each projected node is a GFFN node with inhomogeneous inputs; projected CNNs therefore inherit FFN techniques that do not need homogeneous inputs. Experiments on three ImageNet-pretrained backbones show projection is competitive with PEFT baselines and improves later full fine-tun

What carries the argument

The tensor dot product ⊙_t, which sums channel-wise scalar multiplications of tensors, unifies the two families: GFFN node (Definition 3.3) and GCNN node (Definition 3.4). The proof engine is the bijection between GFFNs and GCNNs whose kernels are size one in every dimension. For the reverse direction, the load-bearing property is separability by input (Definition 3.8): a node's pre-activation decomposes into a sum of sub-functions each depending on one input channel plus a constant bias tensor. Projection (Definition 3.9) rewrites a separable node as σ(Σ γ_{jk} f'_{jk}(Z_k) + b), freezing the f'_{jk} and training the scalars γ_{jk}. This is what turns a projected CNN node into a GFFN node w

Load-bearing premise

The empirical comparison assumes that a fixed simple training recipe — Adam defaults, 20 epochs, no hyperparameter search, and explicitly no data shuffling — is a neutral test of transfer methods; if shuffling and tuning matter for full fine-tuning, the observed advantage of projection may be an artifact of that recipe.

What would settle it

Run the same ImageNet-to-downstream transfer benchmarks with data shuffling enabled and a modest hyperparameter search for each method. If full fine-tuning then matches or exceeds model projection on the same datasets and backbones, the paper's empirical claim that projection is the stronger transfer method under simple recipes would fail. On the theory side, exhibiting any property that holds for all GFFN nodes but fails for some GCNN with 1×1 kernels would contradict Theorem 3.5.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Any result that holds for all GCNNs now applies to all GFFNs, giving a rigorous one-way inheritance from convolutional to feedforward models.
  • Projected CNNs can use FFN training techniques — most directly, one scalar weight per input channel — even though standard CNN kernels have many weights per channel.
  • Projection cuts trainable parameters in a k1×k2 convolution from O(d_out d_in k1 k2) to O(d_out d_in) plus biases, while keeping every convolutional layer adaptable.
  • Model projection is a competitive transfer baseline under simple training recipes and a stronger warm start for full fine-tuning than logistic-regression training.
  • Techniques that rely on homogeneous layer inputs do not transfer to projected CNNs; the inheritance is limited to FFN results that hold with inhomogeneous inputs.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The inheritance framework is stated for FFNs and CNNs, but the separable-by-input condition is the only requirement; the same projection could apply to any architecture whose pre-activations decompose over input channels, including transformer and graph layers.
  • Projection preserves frozen spatial filters and learns channel-wise scalars, which is conceptually a learned channel-attention over fixed features; combining it with adapters or low-rank updates is a natural extension the paper does not try.
  • The strict-subset theorem gives a clean way to test inheritance claims: an FFN technique transfers to all CNNs only if it survives restriction to 1×1-kernel GCNNs, which the paper shows is exactly the GFFN subclass.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. This paper proposes a formal node-level framework for generalized feedforward networks (GFFNs) and generalized convolutional networks (GCNNs), proves that GFFNs form a strict subset of GCNNs (Theorem 3.5), and introduces "model projection": freezing pretrained convolutional filters and learning one scalar coefficient per (output, input) channel pair so that each projected node takes the GFFN form with node-specific inputs (Theorem 3.11). The authors argue that projected CNNs can inherit FFN techniques not relying on homogeneous inputs, and they report transfer-learning experiments on seven datasets and three ImageNet-pretrained backbones, comparing projection against logistic regression, full fine-tuning, and several PEFT baselines. The central claims are the inclusion theorem and the practical effectiveness of projection as a transfer-learning initialization.

Significance. The formal language is clean, and the inclusion direction is correct; the paper makes a useful terminological contribution by distinguishing homogeneous and inhomogeneous inputs when discussing inheritance. Model projection is a simple, interpretable, parameter-efficient method and could be a useful baseline. However, the theoretical centerpieces are near-definitional: the projection is defined so that the GFFN form holds, and the paper does not identify or prove any substantive FFN-derived property that transfers beyond the scalar-weight parameterization. The empirical contribution is potentially valuable, but the current evidence is not reliable because baselines are not controlled and no variance is reported. If the experiments are strengthened with matched protocols and error bars, the paper could serve as a solid baseline paper for parameter-efficient transfer learning.

major comments (4)
  1. [§4.2, Table 2] The comparison against PEFT baselines is not controlled. LoRA-C, LoRA-r8/r32, and PaRA numbers are taken from Ding et al. (2024) and Hedegaard et al. (2024) with no description of their training schedules, data augmentation, epochs, or hyperparameters. The paper's own methods use a deliberately simple recipe (Adam defaults, 20 epochs, no shuffling, no validation selection; §4.1). Any observed difference could be explained by protocol; e.g., Model projection + F.T. (96.95) vs LoRA-C (96.59) on CIFAR-10 is well within possible protocol effects. To support the claim that projection is 'competitive' with PEFT baselines, the authors must re-run the baselines under the same codebase and protocol, or at minimum provide per-method hyperparameters and an ablation showing sensitivity to the recipe.
  2. [§4.1, Tables 3–5] All results come from a single run per condition. There are no random seeds, no error bars, and no repeated trials, yet the paper reports rankings (e.g., Table 3: 'Model projection 17/21 1st') and claims projection 'significantly outperformed' LR and FT. With n=1, these differences are not statistically meaningful. Additionally, the protocol explicitly does not shuffle data and absorbs validation sets into training, which is known to bias SGD/Adam optimizers and can explain the reported FT failures (e.g., 'full fine-tuning occasionally failed to learn... particularly for VGG16'). The authors should provide mean ± std over multiple seeds, enable shuffling, and use standard hyperparameters for FT to ensure a fair comparison.
  3. [§3.2, Def. 3.9 / Thm. 3.11] The main theoretical inheritance claim is true by construction. Definition 3.9 defines the projected node as σ(Σ_k γ_jk f'_jk + b), and Theorem 3.11 simply rewrites this as σ(Ẑ_j ⊙_t γ_j + b) after defining Ẑ_jk = f'_jk(Z_k, W_jk). This is a notational restatement; it establishes that the projected node belongs to the GFFN family, but it does not identify any concrete FFN result or technique whose assumptions are satisfied by projected CNNs beyond the scalar-weight parameterization. Since Corollary 3.12 applies to any property that holds for this class, the paper should demonstrate at least one non-trivial inherited property (e.g., a convergence or approximation guarantee) or explicitly limit the claim to a formal class-inclusion statement.
  4. [Appendix A, proof of Thm. 3.5] The proof shows a bijection between GFFN nodes and GCNN nodes with unit kernels, but it does not directly establish that the network-level classes are in strict inclusion. A network is a composition of layers, and a node-level bijection is insufficient unless one also shows that arbitrary compositions of GFFN nodes map to compositions of unit-kernel GCNN nodes (which is true but should be stated). Moreover, 'strict' requires a specific GCNN function that is not expressible as a GFFN (e.g., a single 3×3 convolution on a 2D input, which mixes spatial locations). The authors should add a short network-level argument and a counterexample.
minor comments (5)
  1. [Corollary 3.6] There is a typo: 'GCCNs' should be 'GCNNs'.
  2. [References] The reference for Adam is given as 'Kingma, D. P. Adam: A method for stochastic optimization' and cited as (Kingma, 2015); the proper citation is Kingma & Ba, 2015, and the full reference is missing.
  3. [Definition 3.1] The notation Z ∈ R^{d×D} is misleading because Z is a vector of d tensors each of shape D; consider using Z = (Z_1, ..., Z_d) with Z_k ∈ R^D, or explaining the notational convention explicitly.
  4. [Appendix A, Lemma A.2] The phrase 'associative for scalar multiplication' is unusual; the property needed is that f'_jk commutes with scalar multiplication (homogeneity), not associativity.
  5. [Figures 2–4] The captions do not state the number of runs or seeds; without this information, the curves could be taken from a single trajectory. Please include the number of seeds and, ideally, error bands.

Circularity Check

0 steps flagged

No significant circularity: the two core theorems are direct algebraic identities from the paper's own definitions, and the empirical evaluation is independent.

full rationale

Theorem 3.5 is proven by an explicit bijection in Appendix A: a GCNN node with kernels of size 1 satisfies Z_k * F_jk = Z_k F_jk, reducing to the GFFN tensor-dot form. This is a direct algebraic reduction, not an assumption of the conclusion. Theorem 3.11 rewrites the projected-node definition from Def. 3.9 by setting hat Z_jk = f'_jk(Z_k, W_jk); the displayed chain of equalities shows the projected node equals sigma(hat Z_j tensor-dot gamma_j + b_j Gamma_Dout), which is exactly the GFFN form of Def. 3.3 with inputs hat Z_j. That is a proof by substitution, not a circular import: projection is defined for any separable node, and the theorem establishes the inherited structure. Corollaries 3.6 and 3.12 are only logical consequences of the stated inclusions. The experiments in Section 4 train on separate train/test splits and compare against fitted baselines; no fitted quantity is renamed as a prediction. No self-citations or author-imported uniqueness theorems carry load. The uncontrolled external PEFT baselines in Table 2 are a benchmark-fairness concern, not a circularity. Thus the derivation chain is self-contained.

Axiom & Free-Parameter Ledger

0 free parameters · 5 axioms · 0 invented entities

No constants are fitted to make the theory work; the projection weights γ_jk are learned in the experiments as part of the proposed transfer method. The axioms are mostly the paper's own node-level definitions and standard facts about convolution.

axioms (5)
  • domain assumption A network is a composition of layers, each producing at least two node outputs (Section 2.1).
    The whole formalization is built on this node-level decomposition; it excludes operations that are not decomposable into per-node sums, such as self-attention or normalization inside the node.
  • ad hoc to paper GFFN (Def. 3.3) and GCNN (Def. 3.4) use tensors and constant-tensor biases.
    These definitions are constructed by the authors to make the inclusion and projection results hold; they are not the standard textbook definitions, which assume scalar inputs (Eq. 1) or 2D images (Eq. 2).
  • standard math A convolution with kernel size 1 in all dimensions is equivalent to scalar multiplication (used in the proof of Theorem 3.5).
    This is a standard property of convolution when the kernel is a single scalar.
  • standard math GCNN node functions are separable by input (Lemma A.1).
    Follows from the definition of convolution as a sum over input channels; used to justify that CNN nodes can be projected.
  • domain assumption ImageNet-pretrained backbones and the seven benchmark datasets constitute a valid testbed for parameter-efficient transfer learning (Section 4).
    No justification for the choice of datasets is given beyond standard practice; results may not generalize beyond this setting.

pith-pipeline@v1.3.0-alltime-deepseek · 12966 in / 22457 out tokens · 195226 ms · 2026-08-03T03:58:38.510873+00:00 · methodology

0 comments
read the original abstract

Neural-network techniques are often transferred across architecture families by analogy, but such transfer is valid only when the assumptions required by a technique are preserved. We introduce this idea as inheritance between model classes. Using a unified node-level framework with tensor-valued activations, we prove that generalized feedforward networks (GFFNs) form a strict subset of generalized convolutional networks (GCNNs), so GCNN properties transfer directly to GFFNs. The reverse direction is not automatic: standard CNN nodes use spatial kernels, while FFN nodes use one scalar weight per input contribution. We introduce model projection to recover a restricted reverse inheritance path. Projection freezes each convolutional input-channel sub-function and learns one scalar coefficient for each input-output channel contribution, giving projected CNN nodes the GFFN-style trainable structure of scalar-weighted input recombination. This inherited structure leads naturally to parameter-efficient transfer learning. Across multiple ImageNet-pretrained CNN backbones and downstream image-classification datasets, model projection is competitive with standard and PEFT baselines and provides an effective initialization for subsequent full fine-tuning.

Figures

Figures reproduced from arXiv: 2602.06245 by Jairo Diaz-Rodriguez, Kelly Ramsay, Nicolas Ewen.

Figure 1
Figure 1. Figure 1: (left) Standard CNN Node: Each input channel contributes through multiple learned weights due to the spatial extent of the kernels. (right) Model projection: The node has exactly one trainable weight per input channel. Spatial structure is preserved by fixed sub-functions, while channel interaction is reduced to a linear weighted sum. The resulting computation is structurally identical to an FFN node with … view at source ↗
Figure 2
Figure 2. Figure 2: Select results from the first experiments. The solid lines show the single stage setup, while the dashed lines show the two stage setup. Each row corresponds to a particular convolutional base, organized from oldest to most recent. The columns show the results on CIFAR 10, CIFAR 100, and Oxford flowers respectively. In all charts, the x-axis is epochs, and the y-axis is test accuracy. home made models, and… view at source ↗
Figure 3
Figure 3. Figure 3: Results from the first set of experiments. The blue lines show the performance of model projection. The green and red lines show the performances of the single layer logistic regression, and the full fine tuning respectively. Each column corresponds to a particular convolutional base, organized from oldest to most recent. Each row shows the results on a particular dataset, organized from largest to smalles… view at source ↗
Figure 4
Figure 4. Figure 4: Results from the second set of experiments. The orange lines show the performance of 2-step fine tuning. The pink lines show the performance of 2-step fine tuning using projection in the first step, and the light blue lines represent the performance of the 2-steps using projection in both steps. Each column corresponds to a particular convolutional base, organized from oldest to most recent. Each row shows… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

35 extracted references · 1 linked inside Pith

  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]

    Abadi, M., Agarwal, A., Barham, P., Brevdo, E., Chen, Z., Citro, C., Corrado, G. S., Davis, A., Dean, J., Devin, M., Ghemawat, S., Goodfellow, I., Harp, A., Irving, G., Isard, M., Jia, Y., Jozefowicz, R., Kaiser, L., Kudlur, M., Levenberg, J., Man\' e , D., Monga, R., Moore, S., Murray, D., Olah, C., Schuster, M., Shlens, J., Steiner, B., Sutskever, I., T...

    2015

  3. [3]

    Y., Suk, H., Suo, M., Tillet, P., Wang, E., Wang, X., Wen, W., Zhang, S., Zhao, X., Zhou, K., Zou, R., Mathews, A., Chanan, G., Wu, P., and Chintala, S

    Ansel, J., Yang, E., He, H., Gimelshein, N., Jain, A., Voznesensky, M., Bao, B., Bell, P., Berard, D., Burovski, E., Chauhan, G., Chourdia, A., Constable, W., Desmaison, A., DeVito, Z., Ellison, E., Feng, W., Gong, J., Gschwind, M., Hirsh, B., Huang, S., Kalambarkar, K., Kirsch, L., Lazos, M., Lezcano, M., Liang, Y., Liang, J., Lu, Y., Luk, C., Maher, B.,...

    arXiv 2024

  4. [4]

    Food-101 -- mining discriminative components with random forests

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

    2014

  5. [5]

    Conv-adapter: Exploring parameter efficient transfer learning for convnets

    Chen, H., Tao, R., Zhang, H., Wang, Y., Li, X., Ye, W., Wang, J., Hu, G., and Savvides, M. Conv-adapter: Exploring parameter efficient transfer learning for convnets. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 1551--1561, 2024

    2024

  6. [6]

    Deep learning with python

    Chollet, F. Deep learning with python. 2021

    2021

  7. [7]

    Chollet, F. et al. Keras. https://keras.io, 2015

    2015

  8. [8]

    Imagenet: A large-scale hierarchical image database

    Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., and Fei-Fei, L. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pp.\ 248--255. Ieee, 2009

    2009

  9. [9]

    Lora-c: Parameter-efficient fine-tuning of robust cnn for iot devices

    Ding, C., Cao, X., Xie, J., Fan, L., Wang, S., and Lu, Z. Lora-c: Parameter-efficient fine-tuning of robust cnn for iot devices. arXiv preprint arXiv:2410.16954, 2024

    Pith/arXiv arXiv 2024

  10. [10]

    Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories

    Fei-Fei, L., Fergus, R., and Perona, P. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. Computer Vision and Pattern Recognition Workshop, 2004

    2004

  11. [11]

    Deep learning

    Goodfellow, I., Bengio, Y., and Courville, A. Deep learning. MIT press, 2016

    2016

  12. [12]

    Facing the elephant in the room: Visual prompt tuning or full finetuning? In The Twelfth International Conference on Learning Representations, 2024

    Han, C., Wang, Q., Cui, Y., Wang, W., Huang, L., Qi, S., and Liu, D. Facing the elephant in the room: Visual prompt tuning or full finetuning? In The Twelfth International Conference on Learning Representations, 2024

    2024

  13. [13]

    Hastie, T., Tibshirani, R., and Friedman, J. H. The elements of statistical learning: data mining, inference, and prediction, volume 2. Springer, 2009

    2009

  14. [14]

    Sensitivity-aware visual parameter-efficient fine-tuning

    He, H., Cai, J., Zhang, J., Tao, D., and Zhuang, B. Sensitivity-aware visual parameter-efficient fine-tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp.\ 11825--11835, 2023

    2023

  15. [15]

    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

    2016

  16. [16]

    Structured pruning adapters

    Hedegaard, L., Alok, A., Jose, J., and Iosifidis, A. Structured pruning adapters. Pattern Recognition, 156: 0 110724, 2024. ISSN 0031-3203. doi:https://doi.org/10.1016/j.patcog.2024.110724. URL https://www.sciencedirect.com/science/article/pii/S0031320324004758

    arXiv 2024

  17. [17]

    J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., et al

    Hu, E. J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., et al. Lora: Low-rank adaptation of large language models. ICLR, 1 0 (2): 0 3, 2022

    2022

  18. [18]

    Huang, G., Liu, Z., Van Der Maaten, L., and Weinberger, K. Q. Densely connected convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.\ 4700--4708, 2017

    2017

  19. [19]

    and Szegedy, C

    Ioffe, S. and Szegedy, C. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning, pp.\ 448--456. pmlr, 2015

    2015

  20. [20]

    Visual prompt tuning

    Jia, M., Tang, L., Chen, B.-C., Cardie, C., Belongie, S., Hariharan, B., and Lim, S.-N. Visual prompt tuning. In European conference on computer vision, pp.\ 709--727. Springer, 2022

    2022

  21. [21]

    Novel dataset for fine-grained image categorization

    Khosla, A., Jayadevaprakash, N., Yao, B., and Fei-Fei, L. Novel dataset for fine-grained image categorization. In First Workshop on Fine-Grained Visual Categorization, IEEE Conference on Computer Vision and Pattern Recognition, Colorado Springs, CO, June 2011

    2011

  22. [22]

    Kingma, D. P. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR 2015), 2015

    2015

  23. [23]

    Kornblith, S., Shlens, J., and Le, Q. V. Do better imagenet models transfer better? In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 2661--2671, 2019

    2019

  24. [24]

    Learning multiple layers of features from tiny images

    Krizhevsky, A. Learning multiple layers of features from tiny images. Technical report, 2009

    2009

  25. [25]

    Convolutional networks and applications in vision

    LeCun, Y., Kavukcuoglu, K., and Farabet, C. Convolutional networks and applications in vision. In Proceedings of 2010 IEEE international symposium on circuits and systems, pp.\ 253--256. IEEE, 2010

    2010

  26. [26]

    Scaling & shifting your features: A new baseline for efficient model tuning

    Lian, D., Zhou, D., Feng, J., and Wang, X. Scaling & shifting your features: A new baseline for efficient model tuning. Advances in Neural Information Processing Systems, 35: 0 109--123, 2022

    2022

  27. [27]

    Convolutional neural networks do work with pre-defined filters

    Linse, C., Barth, E., and Martinetz, T. Convolutional neural networks do work with pre-defined filters. In 2023 International Joint Conference on Neural Networks (IJCNN), pp.\ 1--8. IEEE, 2023

    2023

  28. [28]

    Luo, G. et al. Repadapter: Parameter-efficient and computationally friendly adapter for giant vision models, 2023

    2023

  29. [29]

    Ma, X., Sardy, S., Hengartner, N., Bobenko, N., and Lin, Y. T. A phase transition for finding needles in nonlinear haystacks with lasso artificial neural networks. Statistics and Computing, 32 0 (6): 0 99, 2022

    2022

  30. [30]

    and Zisserman, A

    Nilsback, M.-E. and Zisserman, A. Automated flower classification over a large number of classes. In Proceedings of the Indian Conference on Computer Vision, Graphics and Image Processing, Dec 2008

    2008

  31. [31]

    M., Vedaldi, A., Zisserman, A., and Jawahar, C

    Parkhi, O. M., Vedaldi, A., Zisserman, A., and Jawahar, C. V. Cats and dogs. In IEEE Conference on Computer Vision and Pattern Recognition, 2012

    2012

  32. [32]

    Learning multiple visual domains with residual adapters

    Rebuffi, S.-A., Bilen, H., and Vedaldi, A. Learning multiple visual domains with residual adapters. Advances in neural information processing systems, 30, 2017

    2017

  33. [33]

    and Zisserman, A

    Simonyan, K. and Zisserman, A. Very deep convolutional networks for large-scale image recognition. In 3rd International Conference on Learning Representations (ICLR 2015). Computational and Biological Learning Society, 2015

    2015

  34. [34]

    A real-world dataset and benchmark for foundation model adaptation in medical image classification

    Wang, D., Wang, X., Wang, L., Li, M., Da, Q., Liu, X., Gao, X., Shen, J., He, J., Shen, T., et al. A real-world dataset and benchmark for foundation model adaptation in medical image classification. Scientific Data, 10 0 (1): 0 574, 2023

    2023

  35. [35]

    B., Goldberg, Y., and Ravfogel, S

    Zaken, E. B., Goldberg, Y., and Ravfogel, S. Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pp.\ 1--9, 2022

    2022