acceptodds
Under review as a conference paper at ICLR 2027

Neural compatibility: functional similarity with learning

Abstract

An ongoing problem for computer science and neuroscience is to chart the similarities and differences between how both artificial and biological neural systems represent data and process information. Representational similarity and representational alignment methods seek to quantify and close these gaps. A recent idea is to quantify whether representations from one model are usable by another model by “stitching” activity from one into the other. However, a fundamental property of neural systems is that they learn and adapt to new inputs. We therefore ask what makes two neural systems "compatible" in the sense that one model can efficiently **learn to use** representations from the other. We illustrate with simple examples how "correlation", "stitchability", and "compatibility" between models are in principle distinct ways of conceptualizing similarity between models and brains. Compatibility provides more evidence that representations are most similar across corresponding fractional depths. Despite the shown in-principle distinction, we find empirically that among pretrained image classification models, stitchability and compatibility are strongly correlated. We conclude that 'neural compatibility' is conceptually distinct from other notions of representational alignment, but identifying in-practice impactful differences requires further investigation.

Primary area:
applications to neuroscience & cognitive science
Keywords:
representational similarity, learnability, functional similarity, stitching, neural prosthetics, representational geometry

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_6Q1BOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces the concept of Neural Compatibility to evaluate relationships between neural representations. The authors argue that existing representation similarity measures (e.g., CKA, RSA) and model stitching methods mainly focus on whether representations are similar or directly usable by another model, while overlooking whether a downstream system can effectively utilize the representation after learning and adaptation. The paper defines compatibility as the performance after stitching followed by further fine-tuning, and provides theoretical examples showing that activity correlation, stitchability, and compatibility can produce different rankings of representations. Experiments on ResNet models further investigate the relationship between stitchability and compatibility.

Critical strengths
  1. The paper raises an interesting question: whether neural representations should be evaluated based on their adaptability, rather than only similarity or immediate usability. This perspective is potentially valuable for understanding information transfer, modular neural networks, and representation reuse.

  2. The experiments attempt to validate the existence of compatibility in real neural networks and analyze the relationship between stitching and adaptation, providing initial observations for studying representation transfer and adaptability.

Critical weaknesses
  1. The motivation for introducing compatibility is not sufficiently differentiated from existing concepts such as transferability, fine-tuning efficiency, and representation transfer. Since compatibility essentially measures whether a representation can be effectively utilized after adaptation, its unique scientific value requires further justification.

  2. The definition of compatibility heavily depends on the adaptation process, including the optimizer, learning rate, number of training steps, trainable parameters, and data size. Therefore, compatibility may not be an intrinsic property of a representation, but rather a property determined jointly by the representation, learning algorithm, and adaptation budget. The paper does not provide a sufficiently clear and stable definition.

  3. The theoretical examples mainly demonstrate that different metrics can produce different rankings, but they are highly constructed and do not show whether such differences commonly occur in real neural networks. Moreover, the experiments suggest that ranking reversals between stitchability and compatibility are relatively rare, raising questions about the practical importance of compatibility.

  4. The experiments are mainly conducted on ImageNet-pretrained ResNet models, which limits the generality of the conclusions. It remains unclear whether compatibility can extend to more complex foundation models, such as Transformers, large language models, multimodal models, or biological neural representations.

Questions for authors
  1. What is the fundamental difference between compatibility and existing notions of transferability or fine-tuning performance? If compatibility ultimately measures adapted task performance, why should it be considered a new concept rather than a form of transfer evaluation?

  2. Since compatibility depends on the specific adaptation strategy and budget, how should compatibility be fairly compared across different representations?

  3. Why are ranking reversals between stitchability and compatibility relatively uncommon in real ResNet experiments? If compatibility captures an important new property, why is this phenomenon not observed more frequently?

  4. Can compatibility be extended to unsupervised learning, self-supervised learning, or adaptation processes in biological neural systems? Is the current supervised fine-tuning based definition too restrictive?

1 comment
AuthorsOpenReview ↗

Rebuttal Reviewer 4

W1: Addressed in general response.

W2: We believe it is essential that the measurement is explicit about the learning algorithm and budget. The sensitivity of the measurement ensures conclusions are drawn with learning algorithms and budget in mind similar to how a researcher should consider the implications of the chosen decoder family when comparing representations with decoder probing.

W3: Addressed in general response.

W4: Addressed in general response.

The answers to your questions:

What is the fundamental difference between compatibility and existing notions of transferability or fine-tuning performance? If compatibility ultimately measures adapted task performance, why should it be considered a new concept rather than a form of transfer evaluation? [answer]: Transfer learning typically involves just the upstream parts of a model transferred to a new task or domain, retraining its downstream components from scratch or fine-tuning them. In our notation, this would be a question of the compatibility between a model and its own downstream component, or a model and a randomly-initialized downstream-model. Our notation of ‘compatibility’ focuses on the models themselves and takes pretrained components for the upstream and downstream (motivated by neural prosthetics and stitching) which sets it apart. Since compatibility depends on the specific adaptation strategy and budget, how should compatibility be fairly compared across different representations? [answer]: The adaptation strategy and budget should factor into the analysis from the compatibility measure. However, on a small budget learning rate could have an effect. It is important to ensure that the downstream model receives a decent learning rate from a learning rate finder strategy in such cases. Why are ranking reversals between stitchability and compatibility relatively uncommon in real ResNet experiments? If compatibility captures an important new property, why is this phenomenon not observed more frequently?

[answers:] Our results so far have shown that rank order violations are rare between ResNets, likely because ResNets have highly similar block structures that are repeated over skip connections and share a similar training history. We theorize that more architectural differences between models will yield more variation in neural compatibility, since their mechanisms for feature extraction influence representational format. are more inclined to specific features. For example, position information is implicit in CNN architectures while injected in ViT architectures.

Can compatibility be extended to unsupervised learning, self-supervised learning, or adaptation processes in biological neural systems? Is the current supervised fine-tuning based definition too restrictive? [answers:] We were initially invested in pre-trained architectural differences and did not consider expanding the kinds of finetuning for neural compatibility. Our current results on ResNets as well as Bansal et al (2021)’s model stitching results on self-supervised learning would suggest that rank order violations would also be rare, at least for ResNets. However, Bansal et al (2021) did not do all combinations of possible stitching points which leaves room for further investigation under neural compatibility. It is also an interesting engineering problem for deciding what are the points of stitchability. Does the teacher encoder need to be stitched in? What tasks should be chosen to evaluate the new hybrid models embedding? How are spikes to be translated over a stitching layer? Many interesting questions such as these would need to be answered before expanding.

Yamini Bansal, Preetum Nakkiran, and Boaz Barak. Revisiting model stitching to compare neural representations. In M. Ranzato, A. Beygelzimer, Y. Dauphin, P.S. Liang, and J. Wortman Vaughan (eds.), Advances in Neural Information Processing Systems, volume 34, pp. 225–236. Curran Associates, Inc., 2021. URL https://proceedings.neurips.cc/paper_files/paper/2021/file/01ded4259d101feb739b06c399e9cd9c-Paper.pdf

Reviewer_NXSmOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Neural Compatibility, a measure of functional similarity that considers downstream learning. Unlike traditional model stitching, it allows the downstream network to adapt after connecting representations from different models. The authors provide theoretical examples showing that stitchability and compatibility can produce different model rankings. Experiments on ImageNet-pretrained ResNets further investigate this distinction.

Critical strengths

S1. Interesting research perspective. The paper considers downstream learning when comparing neural representations. This provides a useful extension to traditional representation similarity and model stitching.

S2. Theoretical analysis and empirical validation. The theoretical examples demonstrate possible ranking differences between stitchability and compatibility. The experiments further examine this phenomenon in pretrained neural networks.

Critical weaknesses

W1. Limited methodological novelty and unclear definition. Neural Compatibility is essentially model stitching followed by downstream fine-tuning. The methodological novelty appears limited. Moreover, compatibility depends on the learning algorithm, training budget, and downstream model capacity. It is unclear whether the proposed measure reflects representation compatibility or simply the outcome of a particular training procedure.

W2. Limited empirical evidence. The experiments mainly use closely related ResNet architectures. Only 5 out of 40 downstream configurations show significant ranking reversals. The observed differences are generally modest. These results do not sufficiently demonstrate the practical advantages of Neural Compatibility over traditional stitchability.

Questions for authors

Q1. Beyond adding downstream fine-tuning to model stitching, what is the key methodological contribution of Neural Compatibility?

Q2. How sensitive are compatibility rankings to the training budget, optimizer, and downstream model capacity? Do the rankings remain consistent under different training settings?

Q3. Could the authors evaluate more diverse architectures, such as CNNs and Vision Transformers? Can Neural Compatibility provide meaningful advantages over stitchability in realistic applications?

1 comment
AuthorsOpenReview ↗

Rebuttal Reviewer 3

We appreciate the reviewer for comments on our work and engagement.

W1: We believe we address this in our general rebuttal.

These results do not sufficiently demonstrate the practical advantages of Neural Compatibility over traditional stitchability. W2: Neural compatibility should not be thought of as a replacement for model stitching. Neural compatibility is not a more sensitive and fine-grained measure of representational similarity, it is an expansion to stitching that enables answering new questions. It can be used to answer clinical questions such as if a patient would prefer a prosthetic that requires less learning, but has a lower performance ceiling or a more effective prosthetic which requires more training. It could also be used to answer questions about representational format such as if certain formats that arise from architectural mechanisms are learnable by other models.

Answers to questions:

Q1. Beyond adding downstream fine-tuning to model stitching, what is the key methodological contribution of Neural Compatibility? Can Neural Compatibility provide meaningful advantages over stitchability in realistic applications?

[answer]: Neural compatibility enables answering questions that could not be done with model stitching alone such as clinical questions or questions about representational format. Neural stitching and neural compatibility together could answer a question about how much a model can learn from a representation that is is not trivial reformatting (i.e. rotation). Q2. How sensitive are compatibility rankings to the training budget, optimizer, and downstream model capacity? Do the rankings remain consistent under different training settings? [answers:] Compatibility is directly sensitive to downstream model capacity as that would constrain the representational formats the downstream model would be capable of learning from. A linear downstream would never learn the entirety of a representational format that has a non-linear format. However, neural compatibility could also be meaningful under models with the same capacity. If one model has high neural compatibility on both small and large budgets while the other has high compatibility on only a large budget, it indicates that the first model is similar enough to upstream that it needed less changes to its operations. The sensitivity to capacity and budget help in analysis. We suspect that sensitivity to the optimizer is possible, but we have no results. Rankings should become consistent if all downstream models are given a large enough budget, however, on a small enough budget decisions that affect initial exploration such as learning rate or random initialization may introduce inconsistency.
Q3. Could the authors evaluate more diverse architectures, such as CNNs and Vision Transformers?

[answer] We planned on measuring model stitching and neural compatibility across more diverse neural architectures. However, model stitching has been historically limited to ResNet architectures since attempting to stitch different architectures introduces a new engineering problem. The problem of designing a neural stitching layer that effectively translates representations from one neural model to another without being too powerful.

Reviewer_aiuzOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a relaxed variant of model stitching, which they call "compatibility", to measure similarity of neural representations between two models. Compatibility measures how well one model can learn to use representations from another model, in contrast to stitching which does not allow the downstream model to learn; only a lightweight stitching layer is optimized. The authors provide constructions of when activity correlation methods (like CKA), stitchability, and compatibility differ qualitatively, specifically when they exhibit rank-order differences. Then they run experiments comparing stitching to fine-tuning the downstream model for pairs of ResNet models (3 sizes) on ImageNet. They find that hybrid models along the balanced diagonal already do very well with stitching and that fine-tuning does not help much, but there exist hybrid models off-diagonal where fine-tuning does much better. They also show some pairs exhibit rank-order differences between stitching and compatibility but the results are weak.

Critical strengths

The constructions for where the 3 similarity metrics differ nicely distill the main insight into simple toy examples. It's a clear argument for the existence of theoretical differences between these metrics.

Also the writing is exceptionally clear, thank you for making the paper easy to read.

Critical weaknesses

Unfortunately I think the empirical results as-is are insufficient for an ICLR acceptance, because the primary contribution is toy examples of the differences between metrics, but I don't think the insights are particularly deep. With some additional experiments and discussion I think the paper could be made much stronger.

The main weakness is already highlighted by the authors: the empirical results are very weak in terms of showing a difference between stitching and compatibility. In the abstract they claim "among pretrained image classification models, stitchability and compatibility are strongly correlated" but the only experiments are on ResNets that are highly similar. (I didn't find details on the training setup, maybe I missed it, but I assume the training recipe -- dataset, loss, etc -- is the same for all these models and basically they just differ in size.) not only is this not an ideal setup to substantiate the claim that compatibility can provide novel insights, but also I think it's overclaiming to summarize this result in the abstract as "among pretrained image classification models". I would really want to see results stitching across different losses (e.g. supervised, unsupervised), datasets, etc to support that general of a claim, since I'm not sure how true it is, maybe stitching is modest but compatibility is much better among more different image models.

Separate from empirical results, I would also have liked to see more discussion on where we might expect models in practice to exhibit rank-order differences between stitching and compatibility. The toy example is nice but a bit too stylized and removed from practice.

Questions for authors
  • in sec 3 third paragraph did you mean "best match the original activations from C_upstream" (not C_downstream") or have I misunderstood?
  • can you provide more details on the resnet models in section 3?
  • in the heatmaps, are the results are training or validation or test loss? can you also provide accuracy #s? and also the performance (in loss and accuracy) for the original models.
  • in figure 8 in the appendix I see some examples of non-monotonicity in the fine-tuning loss, e.g. second row second column, first row in the plot. do you have theories why this is?
  • can you provide details on the error bars in figure 4? they look statsig yet the individual runs look extremely noisy.
  • do you have any hypotheses for what is driving the rank-order differences in the experiments?
  • do any hybrids do better than the weaker model in the pair for compatibility? Bansal et al 2021 showed "We also give evidence for the intuition that “more is better” by showing that representations learnt with (1) more data, (2) bigger width, or (3) more training time can be “plugged in” to weaker models to improve performance." I'm curious if you see the same effect with compatibility but for more pairs, say? Maybe these raw results are already in the loss heatmaps but it'd be easier to see if you showed the performance for the original models and delta of hybrid vs weaker model.
  • As I understand it one of the main conclusions from Bansal et al 2021 is that stitching is surprisingly strong due to the Anna Karenina hypothesis, made concrete perhaps by linear mode connectivity. If this is true, (and indeed their results show that stitching recovers performance in many cases quite well), why is fine-tuning the downstream model necessary? In what settings might you expect linear mode connectivity to not hold?
  • activity correlation measures similarity directly, whereas stitching and compatibility measure task performance, or similarity insofar as they are suitable for the same task. is this distinction important at all? if in a hybrid model the downstream model ends up using completely different circuits than it did in-distribution (with its self upstream model), would you consider that similar or different? or say if the hybrid model achieved the same aggregate performance as the original, but behaves totally differently on a per-input basis?
2 comments
AuthorsOpenReview ↗

Rebuttal Reviewer 2

We thank the reviewer for their careful reading, engagement, and many good suggestions.

I would really want to see results stitching across different losses (e.g. supervised, unsupervised), datasets, etc to support that general of a claim, since I'm not sure how true it is, maybe stitching is modest but compatibility is much better among more different image models.

W1: We agree that the rarity of rank-order violations in our experiments likely stems from the inherent similarity of the ResNets we compared; all were downloaded from the torchvision hub of pretrained models and likely have nearly identical training histories. We are interested in expanding our methods into self-supervised models as their training method and encoder structure likely influence representational format.

I would also have liked to see more discussion on where we might expect models in practice to exhibit rank-order differences between stitching and compatibility W2: We theorize that more architectural differences between models will yield more variation in neural compatibility, since their mechanisms for feature extraction constrain the possible representation formats. For example, ViT and CNNs have strictly different formats for location information since ViTs have to inject position embeddings.

in sec 3 third paragraph did you mean "best match the original activations from C_upstream" (not C_downstream") or have I misunderstood?

[answer]: The best match is the original activations of the C_upstream.

can you provide more details on the ResNet models in section 3?

[answer]: We selected three pretrained ResNet models of different ResNet architectures from torchvision hub, all trained on imagenet classification. ResNet18 is a network made of 8 residual blocks. ResNet34 and ResNet50 are made of 16 residual blocks, but ResNet50 has additional convolution layers in its residual blocks.

in the heatmaps, are the results are training or validation or test loss? can you also provide accuracy #s? and also the performance (in loss and accuracy) for the original models.

[answer]: The heatmaps display the validation loss, made from a 10% training-validation split of the data. We also measured the top-1 and top-5 accuracy, as well as measures on the original donor models when appropriate. We will include these in the appendix for easy reference for the future.

AuthorsOpenReview ↗

Rebuttal Question Answers Reviewer 2

Thanks for the attention to detail. Regarding your specific questions:

in figure 8 in the appendix I see some examples of non-monotonicity in the fine-tuning loss, e.g. second row second column, first row in the plot. do you have theories why this is?

[answers]: We identified these cases of non-monoticity as convergence criteria failures because the non-monoticity was caused by decreases in performance and not cases of rank-order violations. An additional interpretation maybe that the learning budget had been too small for these cases.

can you provide details on the error bars in figure 4? they look statsig yet the individual runs look extremely noisy.

[answers]: The line is the validation cross entropy loss of a hybrid model across stitching steps and finetuning steps. The error bar is the variance of the cross-entropy loss across those steps.

do you have any hypotheses for what is driving the rank-order differences in the experiments?

[answers]: We did not identify a consistent pattern within the rank-order differences that would help narrow down the possible hypotheses. The rank-order violation cases with the ResNet34 downstream show to have small amounts of stitching steps, so we suspect that those are cases in which the convergence had been premature and not caused by representational format. In the other cases, the stitching had been effective, but the distance between validation loss had been too small to confidently believe in a representational format difference over noise.

do any hybrids do better than the weaker model in the pair for compatibility? Bansal et al 2021 showed "We also give evidence for the intuition that “more is better” by showing that representations learnt with (1) more data, (2) bigger width, or (3) more training time can be “plugged in” to weaker models to improve performance." I'm curious if you see the same effect with compatibility but for more pairs, say? Maybe these raw results are already in the loss heatmaps but it'd be easier to see if you showed the performance for the original models and delta of hybrid vs weaker model.

[answers]: In general, we found that models that were made of shallow components did poorly in both stitching and compatibility. The “more is better” pattern can be seen when the upstream is deeper but the downstream remains shallow. Uniquely in our experiments, we saw that if the downstream is deep then the neural compatibility will be high despite the upstream component, this is likely due to the skip connections which ensure that all information starting from the input is available for the downstream to learn from. activity correlation measures similarity directly, whereas stitching and compatibility measure task performance, or similarity insofar as they are suitable for the same task. Is this distinction important at all? [answers] Stitching and compatibility measure similarity by interpreting hybrid model behavior. The framework we constructed uses task performance as a way of quantifying paper to enable further analysis and clear comparisons. In principle, another quantification of behavior could be applied.

if in a hybrid model the downstream model ends up using completely different circuits than it did in-distribution (with itself upstream model), would you consider that similar or different? or say if the hybrid model achieved the same aggregate performance as the original, but behaves totally differently on a per-input basis?

[answers]: This is a limitation of all similarity techniques, including activity correlation (Theiss et al, 2026; Golan et al, 2020), as their measures could be driven by which input distributions they were exposed to, despite their underlying circuitry being different. It seems more beneficial to use multiple measurements to investigate representations in the future.

Theiss, M., Braun, L., Saxe, A. M., & Grant, E. (2026). Parameter symmetries determine representational geometry in overparameterized nonlinear networks. arXiv preprint arXiv:2609.39078.

Tal Golan, Prashant C. Raju, and Nikolaus Kriegeskorte. Controversial stimuli: Pitting neural networks against each other as models of human cognition. Proceedings of the National Academy of Sciences, 117(47):29330–29337, 2020. doi: 10.1073/pnas.1912334117. https://www.pnas.org/doi/abs/10.1073/pnas.1912334117

Lukas Braun, Erin Grant, and Andrew M. Saxe. Not all solutions are created equal: An analytical dissociation of functional and representational similarity in deep linear neural networks, 2026. URL https://arxiv.org/abs/2609.38998.

Reviewer_w8tgOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The authors use a stitching method to evaluate when/whether a downstream system can (best) use a given representation. For this, they train an affine stitch, and finetune the model. The authors then test this approach with toy models.

Critical strengths

The paper is quite pedgogical. It clearly separates representational similarity, immediate usability by a frozen downstream model, and usability after downstream learning, with minimal examples that make the distinction intuitive.

The writing is honest: when evaluating the many ResNet upstream–downstream combinations, the authors appropriately acknowledge that ranking reversals are rare and modest.

Critical weaknesses

Contribution:

  • Neural compatibility is not a new method or a new phenomenon, rather it is transferability after adaptation, here reformulated for model stitching. Consequently, the paper renames and formalises a question largely explored in transferability and usable-information probing.

  • The paper is limited to toy examples: e.g. the key example relies on an XOR-like representation, an affine stitch, and a downstream model initialised to behave linearly. These elements mainly highlight that a nonlinear model can learn things that a frozen linear readout cannot. Similarly, the tested models are quite similar: all are pretrained ResNets trained on the same task and data.

Questions for authors
  • What does neural compatibility add beyond classic transferability estimation and usable-information probing?

  • Do the ranking reversals generalize to more advanced settings (SOTA models, large datasets)?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.