Efficient Equivariant Transfer Learning from Pretrained Models

Efficient transfer learning algorithms are key to the success of foundation models on diverse downstream tasks even with limited data. Recent works of Basu et al. (2023) and Kaba et al. (2022) propose group averaging (equitune) and optimization-based methods, respectively, over features from group-transformed inputs to obtain equivariant outputs from non-equivariant neural networks. While Kaba et al. (2022) are only concerned with training from scratch, we find that equitune performs poorly on equivariant zero-shot tasks despite good finetuning results. We hypothesize that this is because pretrained models provide better quality features for certain transformations than others and simply averaging them is deleterious. Hence, we propose λ-equitune that averages the features using importance weights, λs. These weights are learned directly from the data using a small neural network, leading to excellent zero-shot and finetuned results that outperform equitune. Further, we prove that λ-equitune is equivariant and a universal approximator of equivariant functions. Additionally, we show that the method of Kaba et al. (2022) used with appropriate loss functions, which we call equizero, also gives excellent zero-shot and finetuned performance. Both equitune and equizero are special cases of λ-equitune. To show the simplicity and generality of our method, we validate on a wide range of diverse applications and models such as 1) image classification using CLIP, 2) deep Q-learning, 3) fairness in natural language generation (NLG), 4) compositional generalization in languages, and 5) image classification using pretrained CNNs such as Resnet and Alexnet.

Paper

Similar papers

Peer review

Reviewer 8HWc7/10 · confidence 3/52023-07-05

Summary

The paper introduces $\lambda$-equitune, an innovative method that refines existing strategies for achieving equivariant outputs from non-equivariant neural networks. $\lambda$-equitune employs importance weights for feature averaging, which outperforms the group averaging method equine. The authors also present equizero, another approach enhancing zero-shot and fine-tuned performance. The effectiveness of these methods is validated across diverse applications and models, including image classification, deep Q-learning, and natural language generation.

Strengths

**Originality**: The paper introduces a novel approach to addressing the limitations of existing equivariance methods by incorporating importance weights for feature averaging. This creative combination of ideas contributes to the originality of the work. **Quality**: The research is of high quality, backed by rigorous theoretical justifications and empirical evaluations. The authors provide compelling evidence to support the claims made, ensuring the reliability and robustness of their proposed method. **Clarity**: The paper is exceptionally well-written, presenting complex concepts in a clear and concise manner. The organization of the paper facilitates understanding, and the inclusion of theoretical proofs enhances its clarity. **Significance**: The significance of the paper lies in its broad applicability across multiple domains and models, including image classification, deep Q-learning, and natural language generation. The improved zero-shot and fine-tuned results achieved by $\lambda$-equitune and equizero highlight their potential to advance the field of transfer learning.

Weaknesses

One weakness of the paper is the limited exploration of continuous groups, as the focus is primarily on finite groups. The authors acknowledge the need for further work to extend their methods to continuous groups but do not provide concrete solutions or insights in this regard. This limitation restricts the proposed algorithms' generalizability and applicability to real-world applications involving continuous transformations. Addressing this weakness by offering a more detailed discussion on approaches for handling continuous groups would enhance the paper's relevance and broaden its potential impact.

Questions

(1) This paper mentioned that equizero performs well when good agent loss functions are available for downstream tasks. Can the authors elaborate on the process of selecting or designing these proxy loss functions? How did they ensure that the selected loss functions accurately capture the desired goals of each task? (2) Could the authors discuss the potential implications of their research in real-world applications and any limitations or considerations when deploying the proposed methods in practical scenarios?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

Yes

Reviewer vsSr6/10 · confidence 3/52023-07-06

Summary

This paper proposed an equivariant few-shot learning method from pretrained models, namely λ-equitune, averaging the features using importance weights, i.e. λs. These weights are learned directly from the data using a small neural network, leading to excellent zero-shot and finetuned results that outperform equitune. This work further proves that λ-equitune is equivariant and a universal approximator of equivariant functions, and shows that equitune and equizero (the method of Kaba et al. (2022) used with appropriate loss functions) are special cases of λ-equitune. The authors conduct a series of analyses and experiments, validating the simplicity and generality of the proposed method on a wide range of diverse applications and models.

Strengths

(1). The idea of this work is novel for equivariant few-shot learning from pretrained models. (2). A wide range of diverse applications and experiments validate the claim of the work.

Weaknesses

(1). I think Sub-Section '3.2 Properties' is somewhat confusing and should be analyzed in detail to ensure logical consistency. For instance, what is the connection between Theorem 1 and Definition 1, Theorem 2? Is Theorem 1 the foundation of Definition 1 and Theorem 2? Mathematical tools such as definitions, theorems, lemmas, etc. are used to summarize and abstract the theory of the entire method, requiring detailed descriptions to construct their connections. (2). Using importance weights to the features averaging in λ-Equitune need to be discussed, including its benefits against to existing equivariant finetuning methods. Could learnable λ weights adapt the feature outputs from pretrained models? Is this a key to λ-Equitune obtaining excellent results in few-shot learning?

Questions

In Equation (3), maybe M_G^λ (g,x) is M_G^λ (gx)?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

Limitations are provided in the paper.

Reviewer DNsq7/10 · confidence 3/52023-07-07

Summary

This paper proposes an extension of symmetrization approach (Yarotsky 2018; Puny et al., 2021; Kaba et al., 2022; Basu et al., 2023) for achieving invariance and equivariance to (small and finite) symmetry groups, with a focus on empirical demonstration of zero- and few-shot transfer learning from non-equivariant pretrained architectures for a range of applications involving different group symmetries. The technical contribution that allows zero- and few-shot transfer is the introduction of some score (rank) function denoted lambda that weights all possible transformed inputs, and using it to turn standard group averaging (Yarotsky 2018) into weighted averaging in a way that the equivariance (and universality) of symmetrization is still guaranteed. The key idea is that an appropriate choice of the score function allows symmetrization to work favorably for the underlying pretrained model (by assigning higher weights to "important" group transformations), which can allow few-shot transfer learning. The authors name this approach lambda-equitune. In particular, choosing the score function as an indicator on argmin of some loss function allows for a canonicalizing symmetrization (Kaba et al,., 2022) that empirically allows zero-shot transfer, which the authors name equizero. The authors experimentally demonstrate the proposed algorithm in a range of applications including reinforcement learning, fairness in language models, compositional generalization, and image recognition under 90-degree rotations and flips. Yarotsky et al., Universal approximations of invariant maps by neural networks (2018) Puny et al., Frame Averaging for Invariant and Equivariant Network Design (2021) Kaba et al., Equivariance with Learned Canonicalization Functions (2022) Basu et al., Group Equivariant Fine-Tuning of Pretrained Models (2023)

Strengths

S1. The paper aims to address an important and original problem of few-shot or zero-shot transfer of non-equivariant pre-trained deep neural networks to solve equivariant problems. While equivariant transfer learning from non-equivariant model has been investigated by some prior and concurrent work (Basu et al., 2023; Kim et al., 2023), few-shot or zero-shot transfer has not been demonstrated in literature as far as I know. The methodology is clearly motivated and explained, and I think this offers a nice way to steer the behavior of pretrained models towards equivariance with additional controllability offered by the score function or loss function, as demonstrated in the fairness experiment where fairness jointly with high regard score is achieved. S2. The score function lambda for lamdba-equitune and the loss function for equizero doesn't seem to have to respect group symmetry, which is an advantage as it allows for wider range of choices. While this has been theoretically shown in Kaba et al., 2022, as far as I know this is the first work to empirically utilize the property, since Kaba et al., 2022 proposed but did not experiment with optimization approach. S3. The applications demonstrated with experiments are quite comprehensive, ranging from reinforcement learning to language generation and visual recognition. Also, the experimental results overall seems to support the main claims of the paper. Basu et al., Group Equivariant Fine-Tuning of Pretrained Models (2023) Kim et al., Learning Probabilistic Symmetrization for Architecture Agnostic Equivariance (2023) Kaba et al., Equivariance with Learned Canonicalization Functions (2022)

Weaknesses

W1. One major ambiguity I find in the paper is that, while the title of the paper and some parts in the main text mention few-shot learning, it seems the experiments concern zero-shot learning or fine-tuning with a fair amount of data. This was confusing to me given that fine-tuning is not equivalent to few-shot learning. Am I missing something? W2. The approach is only applicable to small, finite groups, due to the requirement of evaluating the score function (lambda) for all possible transformed inputs. This is in contrast to some concurrent work (Kim et al., 2023) that extends to combinatorial or continuous groups, and can be considered a limitation of the proposed algorithm at the current state. W3. For the fairness in language generation, I think there is a limitation in the proposed algorithm that the considered words (upon which group transformations are defined) are implicitly assumed to be not separated by the tokenizer of the language model. This might not be generally true for modern language models, as a tokenizer can choose to split an equality word into non-equality substrings (a potential example is waitress -> wait + ress). In this case, fairness would not be achieved as expected using vocabulary permutation transformations. Kim et al., Learning Probabilistic Symmetrization for Architecture Agnostic Equivariance (2023)

Questions

Q1. For the experiments mentioned as few-shot (e.g., Figure 10), how many shots are used? Q2. In case of zero-shot learning without parameter updates, what kind of practical advantage could we expect from the universality result (Theorem 2)?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

The authors have addressed potential limitations and negative societal impact in Section 6.

Reviewer TNLv5/10 · confidence 4/52023-07-12

Summary

This paper proposes lambda-equitune based on previous works, which averages the features using learned weights by data and small neural network. This paper provides detailed theoretical analysis to prove that the proposed lambda-equitune is equivariant and a universal approximator of equivariant functions. Diverse experiments are conducted to show the generality of the method.

Strengths

1. Efficient transfer learning from foundation models to downstream tasks is an important tasks. This work make interesting improvements on previous work by introducing important weights trained from data and small neural network. 2. Diverse experiments results are provided, including image classification, deep Q-learning and natural language generation. The proposed lambda-equitune method shows good results on majority of the tasks. 3. Detailed theoretical analysis for equitune is provided. The writing and illustration is clear.

Weaknesses

1. The novelty against the previous work of Basu et al. (2023). Since extra parameters and fintuning process are introduced, the contribution of this work could be further explained. 2. According to figure 1, both original data and transformed data are used for inference of the proposed method, it is very natural to conduct embedding from features or from results, how about these embedding methods compared with the proposed method? 3. The setting of the image classification experiments is a bit naive. Since the paper focuses on efficient transfer learning, there are many more meaningful and realistic transfer learning tasks in computer vision domain rather than flip or rotation of 90 degree. 4. Minors: the format of reference part should be adjusted in Page 12.

Questions

Pease refer to the weaknesses part.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

Yes

Reviewer DNsq2023-08-11

Response to rebuttal

Thank you for the comprehensive response. Overall, I find that the rebuttal clearly addresses most of my concerns. I especially appreciate pointing out that score based canonicalization was only considering argmin while this paper extends to weighted average, and also the added extension to large groups, as well as added explanation on usefulness of universality on zero-shot learning. I also agree that revising the title to "Efficient Equivariant Transfer Learning from Pretrained Models" would resolve my concern on few-shot learning. For the review on continuous groups, I was not intending to make explicit comparison to Kim et al., (2023) -- as the authors pointed out, it has been on arXiv after NeurIPS deadline -- but the review was to point out the limitation of the proposed method on large groups. Now I can see that the issue has been resolved with the extended algorithm of weighted averaging combined with canonicalization. Furthermore, it has been also demonstrated empirically with a compelling performance. Overall, I am happy to raise my score from 5 to 7.

Reviewer 8HWc2023-08-16

I am highly satisfied with the authors' responses; their answers have completely addressed my concerns. I recommend accepting this paper in light of the other reviewers' comments.

Reviewer vsSr2023-08-17

I appreciate the detailed response from the authors, which have sufficiently addressed my concerns. I am happy to accept this paper.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC