A Walsh Hadamard Derived Linear Vector Symbolic Architecture

Vector Symbolic Architectures (VSAs) are one approach to developing Neuro-symbolic AI, where two vectors in $\mathbb{R}^d$ are `bound' together to produce a new vector in the same space. VSAs support the commutativity and associativity of this binding operation, along with an inverse operation, allowing one to construct symbolic-style manipulations over real-valued vectors. Most VSAs were developed before deep learning and automatic differentiation became popular and instead focused on efficacy in hand-designed systems. In this work, we introduce the Hadamard-derived linear Binding (HLB), which is designed to have favorable computational efficiency, and efficacy in classic VSA tasks, and perform well in differentiable systems. Code is available at https://github.com/FutureComputing4AI/Hadamard-derived-Linear-Binding

Paper

References (45)

Scroll for more · 33 remaining

Similar papers

Peer review

Reviewer FSGF7/10 · confidence 3/52024-06-23

Summary

The paper introduces a new Vector Symbolic Architecture (VSA) termed Hadamard-derived Linear Binding (HLB), aimed at enhancing computational efficiency and performance in both classical VSA tasks and deep learning applications. VSAs involve binding two vectors to create a new vector in the same space, supporting symbolic operations like binding and unbinding, which is crucial for neuro-symbolic AI. Traditional VSAs, such as Tensor Product Representation (TPR) and Holographic Reduced Representation (HRR), have limitations like computational complexity or numerical stability issues. In contrast, HLB leverages the Hadamard transform, ensuring linear complexity for binding and addressing these drawbacks.

Strengths

- The paper is well organized and the concepts are clearly explained. - The paper details theoretical foundations, derivation processes, and empirical evaluations demonstrating HLB's efficacy across various tasks. It concludes with implications for both classical symbolic AI and modern deep learning systems.

Weaknesses

I do not observe a clear weakness

Questions

Since I am not familiar with this VSA domain, I'm eager to understand the problem setting in the experiment sections 4.2.1 and 4.2.2. Could the author elaborate with simple example on these tasks?

Rating

7

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

The work has no negative societal impacts.

Reviewer rBtB8/10 · confidence 5/52024-06-26

Summary

This work describes a novel vector-symbolic architecture with linear-complexity binding/unbinding operations which includes the following components: - A novel binding/unbinding scheme based on the Walsh-Hadamard transform, which reduces to element-wise multiplication and division due to the self-inverse properties of the Hadamard matrix, the definition of the binding operation, and the proposed projection step. - Initialization of hypervector entries following a bimodal Gaussian distribution for purposes of numerical stability. - A hypervector projection step which in expectation reduces the amount of noise which is incurred by unbinding a vector from a sum of bindings of pairs of vectors. It also simplifies the final form of the bind/unbind operations. - A correction factor that is applied to augment the similarity score when retrieving from a sum of bound vectors, when the number of terms in the sum is known. The authors experimentally test the effectiveness of their novel VSA scheme in several tasks including pure synthetic VSA tasks, as well as two tasks in which VSA schemes have previously been combined with deep neural networks.

Strengths

- The paper is written clearly. - The experimental results in section 4.2 are convincing. - The noise-reducing projection step is an interesting idea. Although the idea is not novel in itself, the definition of the specific projection step is novel as it is well applied in this novel context of Hadamard-derived binding. - This VSA is highly efficient, including only element-wise operations. - The final HLB bind/unbind operations are trivially simple, but there is a novelty and potentially useful insight in their derivation.

Weaknesses

1) Perhaps, it is worth mentioning and if possible elaborating/comparing with other parameterizable binding schemes such as "Steinberg and Sompolinsky, Associative memory of structured knowledge. Sci Rep 12, 21808 (2022)". This binding scheme is based on a trainable matrix. In fact, this parameterized binding scheme is generic by providing a continuum choice from d (VSA) to d^2 (TPR) for d-dimensional embeddings. 2) Besides the synthetic VSA tasks, out of the studied VSA+deep-learning applications, the XML classification results look very promising and practical (although CSPS looks interesting, it is a pseudo-encryption and hence heuristic). The reviewer was wondering if the proposed HLB binding and unbinding operators could have an edge in other hybrid VSA+deep-learning architectures such as: - [27] where MIMOConv used HRR-based binding and MBAT-based unbinding, and MIMOFormer used MAP-based binding and unbinding - Another neuro-symbolic capability of HRRs was studied in "M. Hersche, et al. A neuro-vector-symbolic architecture for solving Raven’s progressive matrices. Nat Mach Intell 5, 363–375 (2023)". Particularly, fractional power encoding was used to describe a set of arithmetic rules with binding and unbinding for abductive reasoning. - Another closely related work to XML classification in which the number of classes C is larger than dimension (d << C) is "M. Hersche, et al. Factorizers for Distributed Sparse Block Codes. Neurosymbolic Artificial Intelligence, 2024". Would be interesting to see how HLB can handle sparse (block-wise) codes. 3) The other key issue with the paper is flaws in its presentation. Some of them are: - The sentences starting at lines 39 and 40 are fully out of place. - Line 107 mentions an “unburdening” operation, clearly referring to “unbinding”. - Line 197 "HLBuses" needs spacing - In Theorem 3.1, the \cdot multiplication notation was used in place of the \odot for element-wise multiplication used otherwise in the paper. - Properties 3.1 is formulated in a weird way: “Let x sampled from omega holds the following properties…” - In properties 3.1, it seems that an expectation operator is missing in the final property concerning the L2-norm of x sampled from MiND. 4) The blue horizontal strip at the top of Figure 4 is a weird anomaly which was not addressed by the authors.

Questions

- It would be great to elaborate on other trainable biding schemes discussed in (1) weaknesses - If time and compute resources allow, it would be interesting to have more results based on the VSA neuro-symbolic architectures pointed out in (2) weaknesses, especially since the first two works have code repositories publically available. - What is the reason behind the blue horizontal strip at the top of Figure 4? - I did not fully understand what the hyperparameter search procedure for experiments in Section 4.2 was. Did the authors re-use the hyperparameters from the experiments in the referenced papers? - In which scenarios is the discussion in section 3.2 relevant? The proposed factor only scales the similarity values, it does not modify the rankings of similarity scores, so it is not entirely clear to me in which context this matters.

Rating

8

Confidence

5

Soundness

3

Presentation

2

Contribution

3

Limitations

This work has no negative societal impact.

Reviewer 6ktE6/10 · confidence 3/52024-07-08

Summary

The authors propose a new form of Vector Symbolic Architecture (VSA), which leverages the Walsh Hadamard transform for vector binding. The new binding is named Hadamard-derived linear binding (HLB), and it achieves comparable or better performance than existing VSAs when performing classic VSA tasks and combined with deep learning.

Strengths

The new binding method is developed with solid mathematical support. With the Hadamard-derived linear binding, the complexity of binding remains at O(d) while also providing some intriguing properties like the stability of similarity scores under the binding of multiple items. In the experiments, the new VSA is able to achieve better or comparable performance than existing VSAs such as MAP and HRR.

Weaknesses

1. The motivation behind using WHT to derive the binding operation is missing. 2. Section 3 (the part before subsection 3.1) is rather hard to follow due to possible typos and poor organization. 3. It is unclear why the proposed method improves the performance in deep learning related tasks (in section 4.2).

Questions

1. Is there any connection between the performance improvements in section 4.2 and the numerical stability of HLB? 2. Is there any benefit of having a symmetric binding? 3. In Definition 3.1, what does B=B* mean? Could you provide a more detailed proof of Theorem 3.1? 4. What are the differences between 'the proof of theorem 3.1' and equation (5)?

Rating

6

Confidence

3

Soundness

3

Presentation

2

Contribution

3

Limitations

The limitations are not provided in the paper. The main limitation is that the merit of the new VSA is unclear in practice as the baselines only involve VSA.

Reviewer xpT17/10 · confidence 4/52024-07-09

Summary

The paper proposed HLB, a vector symbolic architecture derived from the Hadamard transform, reminiscent of holographic reduced representation, to mitigate the challenges that classical VSAs face in deep learning tasks such as numerical stability. Results show comparable memorization capability from alternative models and improved performance on two practical tasks, CSPS and XML.

Strengths

1. The paper is well-presented and delineates the design principles of HLB in the context of other VSA methods, making the position of the HLB clear. 2. HLB is original in the sense that WH transform has not been leveraged directly to derive a VSA before (to the best of my knowledge); although the main design change from HRR is that the Fourier transform is replaced with the Hadamard transform, subsequent derivations of HDC properties and additional technique (the correction term, and the choice of sampling distribution of the hypervectors) is novel. 3. The paper presents reasonably rigorous (see weakness) proof for its claim and supports it with empirical results. Experiments cover fundamental analytical properties desired from a VSA model as well as some applications of VSA in DL. 4. The paper is significant in the field of vector symbolic architecture as a novel VSA model that may address some of the existing challenges in VSA, particularly the numerical stability problem.

Weaknesses

1. Several typos in the draft (line 87 it’s, 95 four, 150 dervise, 197 HLBuses, 260) 2. Although the key parameter selection is explained in the experimental section, the authors defer the main experimental details to the referenced work. The authors are encouraged to briefly summarize each experiment and how HRR is used. 3. See questions for additional concerns.

Questions

1. Is there a way to compute or estimate $\rho$ (number of bound items) given the vector? Many VSA can estimate via its norm and subsequent processing does not need to actively track the number of bound items in practice. 2. How does HLB compare to other VSAs in computational complexity? 3. Is MAP or MAP-C used in the experiment? They seems to be used interchangeably, but the C (which I assume to stand for “continuous” base vectors as opposed to bipolar) in MAP-C was never explained. 4. Although the VSAs in comparisons are selected well, all three approaches leverage approximate unbinding while HLB leverages exact unbinding. How would HLB compare with alternative approaches in general, especially MAP (with bipolar vector initialization) and FHRR (HRR in the Fourier domain)? Both support exact unbinding (due to vectors having exact inverses).

Rating

7

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

No limitation section is present.

Reviewer 6ktE2024-08-08

I am satisfied with the response from the authors, which answers all my questions. The further clarification of the motivation, mathematical derivation, and experiments have made this paper even stronger. Therefore, I changed the score accordingly.

Authorsrebuttal2024-08-10

Thank you!

We are glad we have satisfied your questions and are very appreciative of the score raise! Please let us know if there is anything else that comes up.

Reviewer rBtB2024-08-11

Thanks for the clarifications. We want to re-emphasize the importance of this work that reduces the computational complexity of binding and unbinding from longstanding log-linear (HRR) to linear (HLB). This reduction will surely have profound implications for the use of VSA, especially when integrated with large deep learning models where binding and unbinding could quickly be a computational bottleneck. Due to these broad impacts, I will increase my score to Strong Accept.

Authorsrebuttal2024-08-12

Thank you for score raise and update!

Thank you for the score raise; we will indeed emphasize this in the paper and are glad we could answer your questions. We also have a small update: In running the original MIMO-nets code, the ETA for replicating their results is 191.5 hours. This is longer than the paper reported, so there may still be issues to resolve on our end or some missing detail. We will still endeavor to determine the discrepancy and explore MIMO-nets and other approaches. Thank you again for your review, and we hope to present this work at the conference!

Authorsrebuttal2024-08-12

Small update, XML with MAP-B

We wanted to update you that the largest XML test compiled with consistent results: MAP-B Delicious-200K: nDCG 44.296, and PSnDCG 6.720, both below the scores of our HLB and not discernably worse/better than MAP-C. This would further support that the initialization difference was not a singularly key-factor in our improved performance.

Reviewer xpT12024-08-13

Thank you for the response. The authors have answered all my questions and provided results that make the paper stronger. I have increased the score accordingly.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC