Improving Compositional Generalization Using Iterated Learning and Simplicial Embeddings

Compositional generalization, the ability of an agent to generalize to unseen combinations of latent factors, is easy for humans but hard for deep neural networks. A line of research in cognitive science has hypothesized a process, ``iterated learning,'' to help explain how human language developed this ability; the theory rests on simultaneous pressures towards compressibility (when an ignorant agent learns from an informed one) and expressivity (when it uses the representation for downstream tasks). Inspired by this process, we propose to improve the compositional generalization of deep networks by using iterated learning on models with simplicial embeddings, which can approximately discretize representations. This approach is further motivated by an analysis of compositionality based on Kolmogorov complexity. We show that this combination of changes improves compositional generalization over other approaches, demonstrating these improvements both on vision tasks with well-understood latent factors and on real molecular graph prediction tasks where the latent structure is unknown.

Paper

References (81)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer 7owJ5/10 · confidence 4/52023-06-13

Summary

This paper proposes an iterated learning method with simplicial embeddings (SEM-IL) for systematic generalization. The method is inspired by iterated learning for humans, encouraging compressibility and expressivity. The empirical experiments show the improvement in vision tasks (known latent factors) and real molecular graph prediction tasks (unknown latent factors). It better portrays the generation factors in vision tasks. It also finds that the success of iterated learning may be caused by discrete messages.

Strengths

- The analysis and experiments show the potential of the SEM-IL framework on systematic generalization problems. - It is based on the cognitive science of iterated learning for humans.

Weaknesses

(1) As mentioned in the conclusion section, it is still unclear why the SEM-IL method enables systematic generalization from a theoretical perspective for deep learning. (2) More experiments will be helpful. The paper would be more convincing with a strict theoretical explanation or comprehensive experiments. (3) As mentioned in line 88, the decomposition into G and O is fundamentally unidentifiable because there can be different ways to define factors that work equally well on training data. So it would be helpful to summarize what additional information SEM-IL provides to identify them. (4) It seems source codes are not provided.

Questions

Please address the points in the weakness section.

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

3 good

Limitations

The limitations are mentioned but not in a dedicated section. It does not have a dedicated section for social impact.

Reviewer 7GSE5/10 · confidence 4/52023-07-04

Summary

This work addresses the challenge deep neural networks face with systematic generalization. Systematic generalization refers to the ability to apply learned concepts to new, unobserved combinations. The authors draw on a cognitive science theory known as "iterated learning", a hypothesized process explaining how human language developed this representation ability. The theory is based on simultaneous pressures towards compressibility and expressivity when transmitting messages. The researchers propose using iterated learning, in combination with deep network models that incorporate simplicial embeddings, to generate nearly discrete messages, thereby improving systematic generalization. The effectiveness of this approach is demonstrated in vision tasks with known latent factors and in molecular graph prediction tasks with unknown latent structures, where it outperforms other methods.

Strengths

1. The paper is well-written. The motivation is clear and the problems are stated clearly. 2. The problem addressed in this work is important, and could be particularly interesting to many communities such as causal learning or domain generalization. 3. The empirical performance looks promising, where the gap between the baseline and the proposed approach is non-negligible.

Weaknesses

1. The method section is not described clearly. In particular, in section 3.1, the authors did not explain why IL can be used to improve the performance. Similarly, in section 3.2, I cannot understand the motivation of adopting SEM. Most of the details are deferred to Appendix B, however, it is still important to briefly explain why IL and SEM can lead to the desired results. 2. The main experiments are done with synthetic dataset such as 3dShapes or dSprites dataset. It would be interesting to see how the proposed problem can address the distribution shift problem in domain generalization benchmark such as DomainBed. 3. The model attempts to learn the minimal feature to perform prediction. This could fail when there is “short-cut” presenting in the dataset. In particular, there might be “easy feature” that is highly correlated with the true target feature, where the model could learn to use those easy features instead in the proposed setting. 4. The main theoretical results in Appendix B.3 are adopted from previous work [56], where they consider a distillation setting. 5. The main motivation is based on compressibility and expressivity. The same concept is also proposed in various previous works such as Variational Information Bottleneck (VIM), supervised contrastive learning, invariant risk minimization (IRM), or works in causal representation learning. Nevertheless, there is no discussion or comparison to previous works.

Questions

1. There are three analysis provided in Appendix B to explain IL. It is a bit confusing why we need so many explanation here. In particular, it is hard to related different explanation, e.g., connecting the Bayesian analysis with the KRR analysis. 2. How do we guarantee the predictor g is a good classifier in IL? Do we also train g in the IL procedure? 3. Why can’t IL alone improve the performance? According to Appendix B, it seems that IL alone can address the proposed problem in representation learning.

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

3 good

Limitations

There is no obvious potential negative social impact of this work.

Reviewer r5b66/10 · confidence 3/52023-07-05

Summary

This paper aims to develop methods to improve systematic generalization. The paper proposes several theoretical criteria related to representation learning to improve systematic generalization. Motivated by these principles, the paper then studies whether two methods, iterative learning (IL) and simplicial embeddings (SEM), and their combination can improve systematic generalization. They evaluate these methods on synthetic vision tasks and several tasks related to modeling molecules. They find that the combination of these two methods leads to the highest performance, on average, across these tasks.

Strengths

* I appreciated the formal treatment of the problem in section 2, the intuition offered in section 3, and the empirical analysis connecting the proposed methods to the various hypotheses in these sections. Otherwise it would have been non-obvious how SEM and IL would be expected to improve systematic generalization. * The paper is well written and does a nice job connecting various bodies of work. While the empirical results for the specific proposed methods are not overly convincing (see weaknesses), I could see this work potentially inspiring new ideas. * Different from prior work on systematic generalization that proposes specialized models with task-specific inductive biases, SEM and IL are relatively general methods that do not necessarily limit the expressiveness of the underlying model.

Weaknesses

* While I liked the intuition-building and "ladder of systematicity", I have two concerns with it. First, Stages 1 to 3 seem to directly follow from prior work, e.g. the information bottleneck principle (to the author's credit, this appears to be acknowledged in the appendix). Second, maybe misunderstanding on my part, but I was not fully satisfied with the definition of stage 4. As informal intuition, it makes sense that we would want to learn representations `z` that are similarly "structured" to the latent factors of the data generating procedure `G`. But it wasn't clear to me which operations are expected to be preserved through the proposed "isomorphism". Additionally, while `G` informally represents "semantic factors" it wasn't clear if this can be translated to a formal constraint on the structure of `G`, which we want `z` to be isomorphic to? I think it's fine to leave this mostly informal, but since the paper seemed to attempt to formalize this, I was left a bit unsatisfied. * It seems like most of the intuition offered for SEM and IL corresponds to stages 1 to 3 of the ladder, which per the first point is closely aligned with criteria proposed by prior work. It wasn't clear why these methods should help achieve stage 4, which seemed more unique to the systematic generalization problem setting. * The proposed methods seem to assume a `z` with fixed dimensionality. While this seems reasonable for one of the motivating examples (e.g. `z` representing color and shape), it is less clear how this could be applied to systematic generalization in natural language, which is referenced multiple times in the text. Prior work on systematic generalization in natural language has often considered dynamically-sized tree-structured latent variables. * The overall effectiveness of SEM and IL is hard to gauge. The experiments show improvements on synthetic vision tasks and 3 tasks related to modeling molecules. However, it's unclear whether the proposed methods are effective on real-world vision or NLP tasks that are more popularly studied within the ML community.

Questions

See weaknesses above.

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

4 excellent

Presentation

3 good

Contribution

2 fair

Limitations

I more explicit discussion of limitations could be useful.

Reviewer eFnA4/10 · confidence 4/52023-07-07

Summary

Iterated learning is hypothesized to help human language develop representation ability. The paper proposes to use iterated learning with deep network models containing simplicial embeddings to obtain approximately discrete messages. They show that this combination of changes improves systematic generalization over other approaches, demonstrating these improvements both on vision tasks with well-understood latent factors and on real molecular graph prediction tasks where the latent structure is unknown.

Strengths

The paper aims to improve the systematic generalization capability of deep neural networks, which is currently an important topic in the field. The writing and figures are clear and quite easy to follow.

Weaknesses

Please see in Questions.

Questions

1. A more detailed explanation for iterated learning, which is currently lacking in the paper, will be appreciated. 2. The intuition of using iterated learning in this paper is merely from results in cognitive science. I have not seen convincing reasons why it is appropriate to incorporate iterated learning in deep neural networks, and what limitations of previous works can be tackled by doing so. The introduction should be rewritten to highlight these points. 3. As illustrated in Figure 3 and Table 2 (and the authors also notified), iterated learning alone may not be enough. This again raises a question: Is iterated learning really necessary for this framework? 4. I have not seen a clear connection between iterated learning and simplicial embedding. As far as I understand, the main theme of the paper is about iterated learning, and simplicial embedding empirically improves the performances further (as mentioned in Section 3.2). That is to say, the combination of iterated learning and simplicial embedding in this paper may come from trial and error rather than a thorough theoretical justification. It would be beneficial if the authors can highlight this connection in their paper. 5. The paper mentions compressibility and expressivity pressures throughout the main text. Can we quantify these pressures to show that the proposed method really helps them emerge through iterated learning? I would happily increase my score if the authors carefully tackle my concerns.

Rating

4: Borderline reject: Technically solid paper where reasons to reject, e.g., limited evaluation, outweigh reasons to accept, e.g., good 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

2 fair

Presentation

3 good

Contribution

2 fair

Limitations

The authors have included some discussions regarding the limitations of their work in the conclusion section.

Reviewer 7GSE2023-08-17

Response

Thanks for the rebuttal and additional results. Nevertheless, I still share similar concerns mentioned by other reviewers. For Q4-6, the definitions and the theory are not completely coherent, and one needs to consider various frameworks, e.g., Bayesian and non-Bayesian theory to understand those stages. Q8 is also not being well-addressed, as the other reviewers also stated. Discrete feature has been well-adopted to improve performance, however, combining IL with it does not means IL is the key to improving performance. Overall, I think some presentations and clarifications could be improved, and I will keep my score.

Authorsrebuttal2023-08-17

Concerns about theory and Q8.

Thanks for the reviewer's feedback. We agree that the definitions and theories are not completely coherent, which weakens the paper's contribution. While we recognize that an integrated and self-contained theoretical framework would undoubtedly bolster the paper's strength, we must admit its complexity (and many widely used methods in deep learning need such theories as well). Nevertheless, we believe considering various frameworks is not necessarily a bad thing: some readers might favor describing the same phenomenon from different perspectives, which can make the claim more persuasive. So we put them in the appendix for readers who are interested in IL. For the concerns about Q8, i.e., proving IL is the key to improving performance. We have ablation studies showing that IL+SEM outperforms IL-only, SEM-only, and baseline. When G is given, we also make a prediction on how the confidence of different message pairs evolves and verify that using experiments. We believe these observations are sufficient to claim that under the experimental setting, IL is the key to improving performance. We would appreciate it if the reviewer could give some suggestions (e.g., experimental designs) on how to prove the role of IL. By the way, the original Q8 is asking why IL alone cannot make improvements, which is different from "why IL is the key to improving the performance".

Area Chair sRjF2023-08-17

Reviewer response needed

Hello Reviewer, The authors have endeavoured to address your comments in their rebuttal. The rebuttal phase is a key part of the NeurIPS review process. I invite you to read and respond to the author's comments as soon as possible, latest tomorrow, to give everyone time to continue and conclude the discussion. Thank you for helping make NeurIPS a great conference for our community.

Area Chair sRjF2023-08-17

Reviewer response needed

Hello Reviewer, The authors have endeavoured to address your comments in their rebuttal. The rebuttal phase is a key part of the NeurIPS review process. I invite you to read and respond to the author's comments as soon as possible, latest tomorrow, to give everyone time to continue and conclude the discussion. Thank you for helping make NeurIPS a great conference for our community.

Area Chair sRjF2023-08-17

Reviewer response needed

Hello Reviewer, The authors have endeavoured to address your comments in their rebuttal. The rebuttal phase is a key part of the NeurIPS review process. I invite you to read and respond to the author's comments as soon as possible, latest tomorrow, to give everyone time to continue and conclude the discussion. Thank you for helping make NeurIPS a great conference for our community.

Reviewer 7owJ2023-08-21

Thank you for the rebuttal. I raised the score.

Reviewer eFnA2023-08-22

I thank the authors for their responses. I have raised my score.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC