Fully Distributed, Flexible Compositional Visual Representations via Soft Tensor Products

Since the inception of the classicalist vs. connectionist debate, it has been argued that the ability to systematically combine symbol-like entities into compositional representations is crucial for human intelligence. In connectionist systems, the field of disentanglement has gained prominence for its ability to produce explicitly compositional representations; however, it relies on a fundamentally symbolic, concatenative representation of compositional structure that clashes with the continuous, distributed foundations of deep learning. To resolve this tension, we extend Smolensky's Tensor Product Representation (TPR) and introduce Soft TPR, a representational form that encodes compositional structure in an inherently distributed, flexible manner, along with Soft TPR Autoencoder, a theoretically-principled architecture designed specifically to learn Soft TPRs. Comprehensive evaluations in the visual representation learning domain demonstrate that the Soft TPR framework consistently outperforms conventional disentanglement alternatives -- achieving state-of-the-art disentanglement, boosting representation learner convergence, and delivering superior sample efficiency and low-sample regime performance in downstream tasks. These findings highlight the promise of a distributed and flexible approach to representing compositional structure by potentially enhancing alignment with the core principles of deep learning over the conventional symbolic approach.

Paper

Similar papers

Peer review

Reviewer xKnt7/10 · confidence 3/52024-06-26

Summary

The authors propose the use of the tensor product to model the interactions between object properties and their values, in contrast to the usual concatenation-based fusion for compositional representations. Extensive experiments on a large variety of image datasets are performed, where performance gains are often significantly higher than those of the baselines considered.

Strengths

- Modeling “multiplicative interactions” has proven to be a powerful mechanism for deep learning [1], but is largely neglected in modern times. This paper proposes an interesting way of fusing compositional representations of object properties and their values through the tensor product (and subsequent summation). Beyond the TPR’s success in this paper, I imagine the insights here through the proposed model form could help the design of more expressive deep learning architectures in future work more generally. - The experiments are exhaustive, the technical exposition clear and crisp, and for the most part, every architectural design and decision is justified in great detail. Overall, the paper is remarkably digestible given the technical sophistication, yet offers many insights. --- - [1]: Jayakumar, Siddhant M. et al. “Multiplicative Interactions and Where to Find Them.” *International Conference on Learning Representations* (2020).

Weaknesses

# [W1] Multiplicatively large dimensionality of the TPR representations & issues scaling to larger settings My main criticism of the paper is the resultant TPR's dimensionality. In particular, the soft TPR representations live in a $D_f\cdot D_r$ dimensional space. Due to the use of the Kronecker product, the dimension of the TPR representations grows multiplicatively with the two terms. Whilst the datasets studied in the paper are relatively simple and only 10 factors of variation are modeled, it seems prudent to acknowledge that the TPR size could grow prohibitively large for increased values of $D_f, D_r$ for more complex datasets outside of controlled settings, where significantly larger number of FoVs exist. In particular, the NeurIPS checklist states: `The authors should discuss the computational efficiency of the proposed algorithms`. For example, I see from [L1008] that $N_R:=10$, and $N_F$ is as high as $106$ for Cars3D. Even in this regime with a very small number of roles, the TPR representations are (presumably?) larger than the concatenation-based representations of the baselines. A comparison of FLOPs/multiply-adds needed for the compared methods would be appreciated to better understand the methods’ drawbacks through the use of the (often computationally expensive) tensor product.

Questions

**[Q1]** As expanded upon in [L933], the “role” embedding matrix $M_{\xi_R}$ is initialized with $n$ orthogonal columns but not trained. We are told that it is “intuitive” that role embeddings do not need to convey semantic information, but I think it would be interesting to see an ablation study on this. In particular: if it is the case that the embedding matrix does not need to learn semantic representations, why not make the simpler choice of initializing $M_{\xi_R}$ as a (truncated) identity matrix (this shares the same property of semi-orthogonality)? The decision not to make this parameter learnable is peculiar, and I would be interested in seeing its impact explored experimentally (for example, through the DCI score as in Table. 7).

Rating

7

Confidence

3

Soundness

4

Presentation

4

Contribution

3

Limitations

Limitations are discussed well throughout (and in further detail throughout the appendix), but an extra discussion of the potential drawbacks of increased computational costs should be made.

Authorsrebuttal2024-08-09

Dear Reviewer xKnt, We are glad you found our additional experiments and ablations insightful. Your feedback on our approach, especially regarding the dimensionality, is greatly valued. Thank you again for your time and thoughtful review. Best, Submission14814 Authors.

Reviewer 93av7/10 · confidence 4/52024-07-07

Summary

This work explores compositional representations -- considered a crucial capability underlying intelligent human behaviour in deep learning systems. It argues that there is an incompatibility between discrete symbolic compositional representations—e.g. as obtained through traditional disentanglement approaches—and the continuous vector spaces underlying deep learning systems. To address this, the authors introduce a novel continuous compositional representation that builds on the Tensor Product Representation (TPR) approach, akin to a soft approximation to the TPR. They introduce a method for learning soft TPRs with weakly supervised data called the Soft TPR Autoencoder, and apply it to visual representation learning. This model demonstrates state of the art disentanglement for representation learning and improved sample efficiency for downstream models using those representations.

Strengths

- importance: compositional representation learning, and compositional generalisation and sample efficiency is an important and relatively under-explored area in deep learning research. By focusing on an approach for representation that aims to be more compatible with deep learning this work makes important contributions to this area. - novelty: the method presented is interesting & new but builds on the well established TPR framework. - clarity: the method is explained well and in enough detail to understand and reproduce it. - evaluation: the paper provides a thorough evaluation of the model, including comparisons with the relevant baselines, convergence rates, and performance in low sample regimes. The better performance on downstream tasks w.r.t. number of training examples is particularly interesting and supports some of the motivations for compositional representation. - interested to see where this work might develop in future work exploring hierarchical compositional representations.

Weaknesses

- motivation for the approach & model could be unpacked a bit more (see questions) - other domains: the authors focus on applying the work to the visual data, outside the typical domain of TPR models. In general, a strong feature, enabling them to tackle the messier world of complex visual data (and weaker super vision) and compare the model to the many disentanglement approaches that have previously been applied in the visual domain. However, it could be interesting to see how the _Soft_ TPR approach compares to traditional TPR in its typical domain of language. - no high impact is shown for downstream utility. the improvements seen for downstream tasks are certainly interesting, particularly in the low data regime, however only two tasks are explored. are there any more good downstream tasks to evaluate the utility of their learned representations? could there even be more speculation as to where future work might really leverage what can be learned with this approach? - need for supervision (albeit weakly): could the work be extended to use different forms of supervision or less/no supervision. noting that their ablation showed the importance of the supervision, perhaps the authors could explore more variants on this ablation, and speculate on extensions of the model that could reduce the need for supervision?

Questions

A couple of questions, both on the motivations of the approach in the paper and how it could be improved. In both cases, these are more questions focused on the conceptual motivations and understanding than what the results show: - In terms of motivating the approach, can you be more specific about how representations obtained with traditional disentanglement approaches (typically continuously-valued themselves) could create an incompatibility with the continuous vector space of deep learning systems, and lead to suboptimal performance for the representation learning and downstream models? Any hypothetical or concrete examples (and/or references discussing this point)? - Relative to TPR, what is the relaxation in Soft TPR supposed to be helping with, and why? The manifold example is given, e.g. in Figure 1c, but seems quite vague. Also discussed is that the data being explained maybe only approximately compositional, but why would TPR fail with this and why would we specifically expect soft TPR to help? Could this be unpacked more, perhaps with some concrete examples.

Rating

7

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

N/A

Authorsrebuttal2024-08-13

Dear Reviewer 93av, We are glad our response addressed many of your concerns. Thank you again for your time and insightful review. We really appreciate your thoughtful feedback, Submission14814 Authors.

Reviewer 6VtP5/10 · confidence 4/52024-07-12

Summary

This paper introduces a novel framework for representation learning known as Soft Tensor Product Representations (Soft TPR), aimed at capturing the compositional structure of data more effectively. The authors propose a continuously-valued compositional representation that contrasts with traditional symbolic methods. The paper makes several key contributions in the realm of representation learning, including the conceptualization of Soft TPR, the Soft TPR Autoencoder, and the demonstration of the benefits of this representation for both representation learning and downstream models.

Strengths

The idea of Soft TPR brings a fresh perspective to the field of representation learning, offering a new way to represent compositional structures in a continuous manner. It provides a thorough theoretical exploration of Soft TPR, complete with detailed mathematical proofs and framework extensions. The paper clearly articulates the differences between Soft TPR and existing methods, holding significant potential for enhancing model interpretability and improving robustness to covariate shift.

Weaknesses

1. No MPI scores for disentanglement (in Table 1) are reported. 2. The construction of the model's loss function involves a multitude of hyperparameters, suggesting that the model might require a complex and intricate tuning process. Compared to other models like VCT, the process of adjusting parameters to achieve optimal results could involve additional complexity and such a requirement might pose challenges in terms of computational resources and time. 3. The reason why *Soft TPR is better than TPR is not fully explored.

Questions

1. what is the dissentanglment performance of TPR? The paper only report the downstream results of TPR in Table 6, is it possible to report the disentangment performance of TPR in Table 1? 2. how many pairs are provided for the weakly supervised? For example, in Table 4, the numbers are the total samples for training, or the labeled sample pairs? 3. VCT uses attention-based operations (cross-attention and self-attention) to extract concept vectors, and reconstruct images from the concept tokens, what is the difference between those attention-based operation and the unbinding module or TPR constructor in this paper?

Rating

5

Confidence

4

Soundness

3

Presentation

3

Contribution

2

Limitations

I don't see any negative societal impact.

Authorsrebuttal2024-08-12

Dear Reviewer 6VtP, We are glad that our response addressed most of your concerns. Thank you again for your time and thoughtful feedback regarding our approach. Best, Submission14814 Authors.

Reviewer xKnt2024-08-08

Thanks to the authors for their thorough response. The rebuttal was a pleasure to read; the extra experiments and ablations are insightful, and the reviewers make a good point that—despite the dimensionality of the space growing multiplicatively—the independent dimensionality of the two factors means this is not always too problematic. The “weaknesses” raised in my initial review were only very minor in the first place—the authors have acknowledged the multiplicative growth of the space, and I see no critical reasons why the paper should not be accepted.

Reviewer 6VtP2024-08-12

Thanks for the response. Most of my concerns are addressed. I have raised my score.

Reviewer 93av2024-08-12

Thanks for the detailed and clear follow-up, many of concerns have been addressed and I have raised my score.

Bethia Sun12025-01-15

Please note there is a small error in the author rebuttal above, under section *1) Incompatibility between disentangled representations and deep learning’s continuous vector spaces*, where the tensor product is mistakenly taken over $\psi_{col}(purple) \otimes \psi_{sh}(square)$ to produce the representation, instead of computing $\xi_{F}(purple) \otimes \xi_{R}(col) + \xi_{F}(square) \otimes \xi_{R}(sh)$, but the conclusion remains exactly the same. (Note that $\psi_{col}(purple)= \xi_{F}(purple) \otimes \xi_{R}(col)$ and $\psi_{sh}(square)= \xi_{F}(square) \otimes \xi_{F}(sh)$ should *both* be 4-dimensional, *not* 2-dimensional).

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC