The challenge in object-based visual reasoning lies in generating concept representations that are both descriptive and distinct. Achieving this in an unsupervised manner requires human users to understand the model's learned concepts and, if necessary, revise incorrect ones. To address this challenge, we introduce the Neural Concept Binder (NCB), a novel framework for deriving both discrete and continuous concept representations, which we refer to as"concept-slot encodings". NCB employs two types of binding:"soft binding", which leverages the recent SysBinder mechanism to obtain object-factor encodings, and subsequent"hard binding", achieved through hierarchical clustering and retrieval-based inference. This enables obtaining expressive, discrete representations from unlabeled images. Moreover, the structured nature of NCB's concept representations allows for intuitive inspection and the straightforward integration of external knowledge, such as human input or insights from other AI models like GPT-4. Additionally, we demonstrate that incorporating the hard binding mechanism preserves model performance while enabling seamless integration into both neural and symbolic modules for complex reasoning tasks. We validate the effectiveness of NCB through evaluations on our newly introduced CLEVR-Sudoku dataset.
Paper
Similar papers
Peer review
Summary
The paper proposes a novel approach to unsupervised concept learning based on both continuous and discrete encodings. Neural Concept Binder (NCB) allows humans inspecting and revising the learnt concepts. In the experiments, NCB’s discrete concept encodings result as expressive as the continuous encodings. Also, NCB can be integrated with symbolic and sub symbolic module. Finally, to support the experimental evaluation the paper introduces a novel dataset CLEVER-Sudoku, very suitable for neuro-symbolic benchmarking.
Strengths
- **Novelty**: the proposed approach, although based on existing works SySBinder and Slot attention, is surely novel in the field of concept learning and potentially very relevant as it may strongly facilitate the extraction and discovery of unsupervised concepts. Particularly the possibility to revise concepts is completely novel to the best of my knowledge and very useful to improve human-computer interaction. - **Novel resource** presented: CLEVER-Sudoku will be surely an important resource for the Neuro-symbolic literature.
Weaknesses
## Major issues: * Method presentation: - The way in which block-slot-encodings are obtained, is badly presented. Although it is based on previous literature, since it is a key architectural component, it should have been presented more in details. I suggest the author to employ a background section to report the way in which slot attention and sys binder work, in order to make the paper self-contained. - Figure 2 which illustrates the core of the method is quite confusing: it is not clear how the discrete concepts are actually represented (the concept slot encodings reported are positive continuos representations). Also, the reminders to the figures in the text do not help as they generically refer to the entire figure and not to a specific block. A color coding of the different parts of the model could help understanding. - How does the $\texttt{enc}_l^j$ work is not clear. What does it receive in input? Where it is extracted from? - All the revision operations are definitely not clear. The formal operation to be executed is often confusing. * Experimental evaluation: - Models: NCB has been compared only against SysBinder. While it is a very novel and innovative method, there is a complete lack of benchmarking against standard unsupervised concept-based approaches such as SENN[1], BotCL[2], ACE[3]. Comparing against supervised approaches such as CBM[4] or CEM[5] could have been also useful. - Datasets: NCB is only tested on variants of CLEVER. While it is surely an interesting benchmark, real-world benchmarks are missing. Experiments on CUB or CELEBA, for instance, would have been very appreciated to better understand the scalability of the approach. ## Minor issues * Related work: - The unsupervised concept learning literature does not review several important concept-based paper working both post-hoc and explainable by design. Some examples are SENN[1], ACE[3], VAEL[6], as well as notorious prototype-based approaches such as Prototype layer [7] and ProtopNets[8]. - Unlike you state, continuous and discrete representations have been combined in recent literature for supervised concept learning. Some examples are CEM[5] and ProbCBM[9]. * Unclear sentences: - “Briefly, given an image x, NCB derives a symbolic representation, c, which expresses the concepts of the objects in the image, i.e., object-factor level concepts. Herefore, NCB infers a block-slot encoding, z, of the image and performs a retrieval-based discretization step to finally infer concept-slot encodings, c”. The consequentiality of the inference process is misleading from this sentence. * Method Inspection. What the authors refer as implicit, comparative, interventional and similarity-based inspections are normally referred to as example-based explanations (implicit and similarity-based) and counterfactual explanations (comparative and interventional). Sticking to well-known terms in literature is a good choice to avoid further confusion in the reader. Overall, I think it's an interesting paper proposing a novel approach to unsupervised concept learning. However, I think it will benefit from a further revision to deeply improve method presentation and expand the experimental campaing including other standard unsupervised concept-learning approaches and datasets. [1] Alvarez Melis, David, and Tommi Jaakkola. "Towards robust interpretability with self-explaining neural networks." Advances in neural information processing systems 31 (2018). [2] Wang, B., Li, L., Nakashima, Y., and Nagahara, H. “Learning bottleneck concepts in image classification”. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023). [3] Ghorbani, Amirata, et al. "Towards automatic concept-based explanations." Advances in neural information processing systems 32 (2019). [4] Koh, Pang Wei, et al. "Concept bottleneck models." International conference on machine learning. PMLR, 2020. [5] Espinosa Zarlenga, Mateo, et al. "Concept embedding models: Beyond the accuracy-explainability trade-off." Advances in Neural Information Processing Systems 35 (2022): 21400-21413. [6] Misino, Eleonora, Giuseppe Marra, and Emanuele Sansone. "Vael: Bridging variational autoencoders and probabilistic logic programming." Advances in Neural Information Processing Systems 35 (2022): 4667-4679. [7] Li, Oscar, et al. "Deep learning for case-based reasoning through prototypes: A neural network that explains its predictions." Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 32. No. 1. 2018. [8] Chen, Chaofan, et al. "This looks like that: deep learning for interpretable image recognition." Advances in neural information processing systems 32 (2019). [9] Kim, Eunji, et al. "Probabilistic Concept Bottleneck Models." International Conference on Machine Learning. PMLR, 2023.
Questions
A few questions to try to understand better the notation employed to define the revision operations. - What do the authors mean with $v_l \rightarrow v_m$? - What does the add operation work and how one can provide an encoding for a concept and be sure the network employs it as intended?
Rating
5
Confidence
4
Soundness
3
Presentation
1
Contribution
3
Limitations
The method limitations are well addressed by the authors.
Additional background section
**Background** The binding mechanism (SysBinder) of Singh et al. (2024) allows images to be encoded into continuous block-slot representations and relies on the recently introduced slot attention mechanism [2]. In slot attention, so-called slots, $s \in R^{N_S \times N_B D_B}$ (each slot has dimension $N_B D_B$), compete for attending to parts of the input via a softmax-based attention. These slot encodings are iteratively updated and allow to capture distinct objects or image components. The result is an attention matrix $A \in R^{N_S \times D}$ for an input $x \in R^{D}$. Each entry $A_{i}$ corresponds to the attention weight of slot $i$ for the input $x$. Based on the attention matrix, the input is processed to read-out each object by multiplying $A$ with the input resulting in a matrix $U \in R^{N_S \times N_B D_B}$. SysBinder now performs an additional factor binding on the vectors $u_i$ of $U$. The goal of this factor binding mechanism is to find a distribution over a codebook memory for each block in $u_i$, i.e., $u_{i}^j$. This codebook memory (one for each block), $M^j \in R^{K \times D_B}$, consists of a set of $K$ learnable codebook vectors. Specifically, for each block $j$ an RNN consisting of a GRU and MLP component iteratively updates the $j$-th block of slot $s_i$, $s_{i}^j$, based on $u_i^j$ and previous $s_{i}^{j}$. Finally, a soft information bottleneck is applied where each block $s_i^j$ performs dot-product attention over the codebook memory leading to the final block-slot representation: $$ \mathbf{s}_{i}^j=\left[\underset{K}{\operatorname{softmax}}\left(\frac{\mathbf{s}_i^j \cdot (\mathbf{M}^j)^T}{\sqrt{D_B}}\right)\right] \cdot \mathbf{M}^j $$ This process is iteratively refined together with the refinement processes of slot attention. Overall, the encodings of SysBinder represent each object in an image by a slot with $N_B$ blocks where each block represents a factor of the object like shape or color. Note that in the main text, the final $s_i^j$ is denoted as $z_i^j$. --- [1] Singh, Gautam, Sungjin Ahn, and Yeongbin Kim. "Neural Systematic Binder." ICLR, 2023. [2] Locatello, Francesco, et al. "Object-centric learning with slot attention." NeurIPS, 2020.
Updated text on merge revision
(i) Merge Concepts: In the case that $\mathcal{R}$ contains multiple concepts that represent a joint underlying concept (e.g., two concepts for purple in Fig.3 (right)) it is easy to update the model's internal representations by replacing the concept symbols of one concept with those of a second concept. Specifically, according to human or additional model feedback, if concept $m$ in block $j$ should be merged with concept $b$ ($m,b \in \{1, \cdots, N_C\}$) then for all corpus tuples, $(\texttt{enc}_l^j, v_l) \in R^j$, we replace $v_l$ with $b$ if $v_l = m$.
Updated text on add revision
(iii) Add Encodings or Concepts: If a specific concept is not sufficiently well captured via the existing encodings in $\mathcal{R}^j$, one can simply add a new encoding for the concept, $m$, to the corpus: $\hat{\texttt{enc}}_{l+1}^j$ This leads to an additional entry in the corpus: $(\hat{\texttt{enc}}_{l+1}^j, m)$ Accordingly, it is also possible to add encodings for an entire concept. Hereby, via the soft binder one infers block encodings of example objects that represent that novel concept, $b$, and adds these to the corpus as $(\hat{\texttt{enc}}_{l+1}^j, b)$ with $b = N_C+1$.
Updated unclear sentence
Briefly, given an image, $x$, NCB infers latent block-slot encodings, $z$, and performs a retrieval-based discretization step on $z$ to infer concept-slot encodings, $c$. These express the concepts of the objects in the image, i.e., object-factor level concepts.
I thank the authors for their efforts in trying to address the suggested issues. However, I will increase my score to 5 only since I still think is a borderline paper: i) The method presentation should have been improved according to what the authors have reported, but I should review it again to assess it better. ii) Testing only on toy datasets is still not acceptable for a neurips paper. Although I agree that better object centric learning will improve the performance of the model, how the current method behave on natural image dataset is of crucial importance to globally assess the model. Bad, but promising results would have still been appreciated.
We thank the reviewer for their time and for reconsidering their rating. However, we disagree that non-synthetic datasets are mandatory for NeurIPS. In fact, several influential and recent NeurIPS papers in the field of object-centric/NeSy learning were published based only on synthetic datasets, e.g., [1,2,3,4]. [1] Locatello, Francesco, et al. "Object-centric learning with slot attention." NeurIPS, 2020. [2] van Krieken, Emile, et al. "A-nesi: A scalable approximate method for probabilistic neurosymbolic inference." NeurIPS, 2023. [3] Marconato, Emanuele, et al. "Not all neuro-symbolic concepts are created equal: Analysis and mitigation of reasoning shortcuts." NeurIPS, 2023. [4] Li, Zenan, et al. "Neuro-symbolic learning yielding logical constraints." NeurIPS, 2023.
Summary
This paper introduces neural concept binder, a neural symbolic framework that utilizes both soft and hard binding. Building on top of the sysbinder model, it can additionally do exemplar-based hard binding and revise concepts. Evaluations made on CLEVR and the proposed CLEVR-Sudoku dataset proved the method's validity.
Strengths
- The paper is well-written and easy to read. Connections to previous works are clarified nicely. - It's good to see the incorporations of both hard and soft bindings to existing neural-symbolic frameworks. - The model achieved good performance on the proposed CLEVR-Sudoku task and can do satisfactory concept revision and inspection, which is a neat proof of concept that hard binding works.
Weaknesses
There are several weaknesses I can foresee that may lead to the rejection of this paper. - Limited contribution: After so many years of developing neural-symbolic methods in visual reasoning, from the earliest modular approaches to unsupervised concept learners, code-based reasoning models, and recent visual programming-like frameworks, the goal of neural-symbolic modeling has dramatically changed. In this work, the neural concept binder still focuses on one of the earliest task categories designed for visual reasoning (CLEVR attribute classifications or unsupervised concept learners). It's also built on top of sysbinder, in other words, it's merely an incremental improvement by adding a retrieval-based library. - I don't see any generalizability of this method beyond extremely toy tasks (attribute classification). The proposed CLEVR-Sudoku is strange and does not correspond to any of the real-world visual reasoning tasks. Relational tasks are also not tackled in this paper.
Questions
- How generalizable can this method be? Can it serve as a part of the closed-loop reasoning in CLEVR (I mean, the original CLEVR questions, as tackled in NS-CL and NS-VQA line of work)? - Can relational concepts be similarly represented via soft/hard binding?
Rating
4
Confidence
5
Soundness
3
Presentation
3
Contribution
1
Limitations
Yes, the authors have adequately addressed the limitations. I do not see any potential negative societal impact.
References
[1] Yi et al. "Neural-symbolic vqa: Disentangling reasoning from vision and language understanding." NeurIPS, 2018. [2] Koh et al. "Concept bottleneck models." ICML, 2020. [3] Shindo et al. "α ILP: thinking visual scenes as differentiable logic programs." Machine Learning, 2023. [4] Surís et al. "Vipergpt: Visual inference via python execution for reasoning." CVPR, 2023. [5] Gupta et al. "Visual programming: Compositional visual reasoning without training." CVPR, 2023. [6] Hsu et al. "What’s left? concept grounding with logic-enhanced foundation models." NeurIPS, 2024. [7] Mao et al. "The neuro-symbolic concept learner: Interpreting scenes, words, and sentences from natural supervision." ICLR, 2019. [8] Wu et al. "Neural Language of Thought Models." ICLR, 2024. [9] Wüst et al. "Pix2Code: Learning to Compose Neural Visual Concepts as Programs." UAI, 2024. [10] Mitchell, Melanie. "Abstraction and analogy‐making in artificial intelligence." Annals of the New York Academy of Sciences, 2021. [11] Zhang et al. "How Far Are We from Intelligent Visual Deductive Reasoning?." ICLR 2024 Workshop: How Far Are We From AGI. [12] Lake et al. "The Omniglot challenge: a 3-year progress report." Current Opinion in Behavioral Sciences, 2019. [13] Chollet, François. "On the measure of intelligence." arXiv, 2019. [14] Moskvichev et al. "The conceptarc benchmark: Evaluating understanding and generalization in the arc domain." arXiv, 2023. [15] Raven, Jean. "Raven progressive matrices." Handbook of nonverbal assessment, 2003. [16] Lake et al. "Human-level concept learning through probabilistic program induction." Science, 2015. [17] Nie et al. "Bongard-logo: A new benchmark for human-level concept learning and reasoning." NeurIPS, 2020.
Thanks for the authors' response. After reading the rebuttal, I've decided to increase the score by 1.
We thank the reviewer for the time! However, we kindly ask the reviewer to raise remaining concerns in case we could not resolve all issues with our initial response.
Summary
The authors introduced a pioneering framework that combines an object-centric learning module with a retrieval-based module to address visual reasoning tasks and a new visual reasoning task, CLEVR Sudoku. The proposed method demonstrated significant potential in effectively acquiring inspectable and revisable concepts via human or machine feedback in various scenarios.
Strengths
- S1: The proposed method offers significant novelty in that it has the potential to serve as a building block for concept learning, which can be leveraged as a core module in other frameworks. The author's well-structured experiments provided compelling evidence in support of these claims.
Weaknesses
- W1: The proposed approach can be interpreted as directly integrating SysBinder and HDBSCAN. Because the initial concept detection fundamentally depends on the complete functionality of SysBinder, this framework may not circumvent specific inherent challenges of object-centric learning, including inductive bias resulting from choosing the proper object-factor encoder and identifiability issues. - W2: Using HDBSCAN is intuitive in the proposed method, but it would be beneficial to include an additional experiment that compares different clustering methods.
Questions
Please check out the Weakness section first. I listed the following questions and suggestions that would be helpful for authors' future works: - Q1: The recent method [1] in object-centric learning literature is linked to causal representation learning and the identifiability of slot representations. How can this be integrated into your framework? - Q2: Object-factor learning can be interpreted as learning atoms in logic, and the NN explanations in Table 2 can be seen as the simple form of propositional logic in a neuro-symbolic framework. How can an object-centric learning framework be extended to represent logical rules, such as in the form of first-order logic? Reference - [1] Mansouri, Amin, et al. "Object-centric architectures enable efficient causal representation learning." arXiv preprint arXiv:2310.19054 (2023).
Rating
7
Confidence
3
Soundness
3
Presentation
4
Contribution
4
Limitations
Please check out the Weakness and Question sections.
Summary
This paper introduces Neural Concept Binder, a framework for obtaining discrete concept representations from images without any supervision. The method is an extension of Neural Systematic Binder (SysBinder), adding a clustering step on top of the block-slot representations to obtain discrete concept representations. The resulting representations are interpretable and modifiable, as shown in the experiments. The model is additionally evaluated on property prediction and downstream tasks on modifications of the CLEVR dataset and shown to be able to leverage fewer training samples than SysBinder.
Strengths
The paper investigates an important problem in learning discrete, interpretable concepts from images in an unsupervised way. The model is a logical extension of SysBinder in clustering the representations to obtain discrete concepts. The experiments show improvements in sample efficiency of these discrete representations over SysBinder’s continuous representations.
Weaknesses
1. Since the solver used in the Sudoku experiments is the same across all baselines, it seems the determining factor of performance is in how well the digits are classified. Therefore, I do not believe framing this evaluation in the context of Sudoku adds any insight—in fact it seems to add unnecessary noise to the evaluation. The evaluation in the appendix (Figure 8) seems more informative and sufficient for determining the benefit of NCB. 2. This paper is missing several related citations: Unsupervised Concept Discovery Mitigates Spurious Correlations (https://arxiv.org/abs/2402.13368) and Neural Language of Thought Models (https://arxiv.org/abs/2402.01203). NLoTM is particularly relevant and can be an additional baseline since it also extends SysBinder to learn discrete representations, except it is trained in an end-to-end way. 3. The discussion in section 3.3 is interesting, but it would be informative to tie each point with corresponding experimental evidence.
Questions
1. For SysBinder (hard) and SysBinder (step), do the models train well with this temperature adjustment? E.g. do they exhibit decent reconstruction and slot decomposition? 2. I’m not sure if I completely understand the analysis for Q4, but since this is done on the concept encodings, and not the discrete representations, can the same analysis be done with SysBinder representations? If so, does NCB offer any additional benefits here? 3. How important is the choice of clustering algorithm to the results? What if we use a simple k-means clustering as is done in the original SysBinder paper?
Rating
6
Confidence
3
Soundness
3
Presentation
2
Contribution
2
Limitations
Yes
Dear Reviewers, As the discussion period is soon approaching, we wish to thank everyone for the feedback provided on our manuscript. In light of the revisions and clarifications made, we believe we have addressed the concerns that were previously raised. Particularly, we would like to highlight the new comparison to the NLOTM baseline (which we have now completed for all seeds; see below) and several ablations on the clustering component of NCB. We are encouraged by the positive feedback and hope that our responses have resolved the reviewer's questions. In this context, we respectfully ask the reviewers to reevaluate their initial score and are happy to answer any remaining questions or clarifications. Thank you once again for your time, expertise, and consideration. ### Remaining results for NLOTM By now, the experiments on the NLOTM baseline are finished for all seeds. As the preliminary results on the first seed indicated, it is substantially harder to infer the ground-truth concepts from NLOTM's representation than NCB's. | N Train | NCB | seed 0 | |---------|---------|--------------| | N=2000 | 99.02+-1.00 | 84.36+-8.54 | | N=200 | 98.50+-1.80 | 72.99+-8.43 | | N=50 | 95.87+-2.93 | 49.94+-4.97 | | N=20 | 94.22+-4.11 | 37.05+-4.11 |
Response to Rebuttal
Thank you for answering my questions and taking the time to run additional experiments. Overall, I think this is a nice extension of SysBinder, although I am still not convinced that the Sudoku experiments are necessary to demonstrate the model's capabilities. I have decided to increase the score.
We thank the reviewer for their time and reconsidering their rating!
Reponse to the author's rebuttal
Thank you for the authors' efforts in providing additional experiments on clustering and the reference. I'm maintaining the score.
We thank the reviewer for their valuable time and appreciation for our work!
Decision
Accept (poster)