Human reasoning can be understood as a cooperation between the intuitive, associative "System-1" and the deliberative, logical "System-2". For existing System-1-like methods in visual activity understanding, it is crucial to integrate System-2 processing to improve explainability, generalization, and data efficiency. One possible path of activity reasoning is building a symbolic system composed of symbols and rules, where one rule connects multiple symbols, implying human knowledge and reasoning abilities. Previous methods have made progress, but are defective with limited symbols from handcraft and limited rules from visual-based annotations, failing to cover the complex patterns of activities and lacking compositional generalization. To overcome the defects, we propose a new symbolic system with two ideal important properties: broad-coverage symbols and rational rules. Collecting massive human knowledge via manual annotations is expensive to instantiate this symbolic system. Instead, we leverage the recent advancement of LLMs (Large Language Models) as an approximation of the two ideal properties, i.e., Symbols from Large Language Models (Symbol-LLM). Then, given an image, visual contents from the images are extracted and checked as symbols and activity semantics are reasoned out based on rules via fuzzy logic calculation. Our method shows superiority in extensive activity understanding tasks. Code and data are available at https://mvig-rhos.com/symbol_llm.
Paper
Similar papers
Peer review
Summary
The paper addresses activity reasoning, where the task is to detect human activities given visual inputs. To achiever compositional generalization for activity understanding, the paper propose a new symbolic system that is capable of relational reasoning. In the experiments, the proposed system improved naive CLIP models showing its strong reasoning capability.
Strengths
Overall, the paper addresses an important problem for integrating symbolic reasoning with neural perception models. The paper demonstrates an instantiation for such an integration by providing formal definitions. The paper conveys its idea with a good presentation using clear visualizations. The obtained results are interesting. These contributions are beneficial for the community to build agents that can perform complex reasoning and understand the world better beyond simple visual perception.
Weaknesses
My major concerns regarding this paper are listed as follows. - The notation and semantics of the proposed method are unclear. The rule is denoted using the logical entailment $\models$. Is this different from implication $\rightarrow$, which is commonly used for logic? In Tab. 2 in the appendix, $\rightarrow$ is used. Since the definition of logical entailment is not provided in the paper, it is unclear how the rules can be understood as formulae. The implication and entailment articulate different concepts. - Related to the first point, it is not clear how the proposed symbolic system is related to other symbolic reasoners, e.g. standard first-order logic? What are the pros and cons? - The paper mainly addresses activity reasoning, but it is not well-described in the paper. Even though some visualizations are provided in the paper, it is not easy to follow the arguments for readers who are not familiar with this topic. It would be beneficial to provide some simple examples throughout the paper. - Related to the previous point, the empirical setting is not easy to follow. E.g. on line 232 , it says, "We follow the setting of HAKE [20] to integrate the reasoning ...", however, this explanation is not self-contained. The experimental settings should be articulated more clearer in the paper. - Other minor issues: - Definition 5 is not clear. What are activity semantic and ongoing symbols?
Questions
- The rule is denoted using the logical entailment $\models$. Is this different from implication $\rightarrow$, which is commonly used for logic? (In Tab. 2 in the appendix, $\rightarrow$ is used.) - How the proposed symbolic system is related to other symbolic reasoners, e.g. standard first-order logic? What are the pros and cons?
Rating
6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.
Confidence
2: You are willing to defend your assessment, but it is quite likely that you did not understand the central 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 listed some limitations in the paper.
Summary
This work is in the area of human visual activity reasoning and tries to improve System-2 (logic, deliberative reasoning) integrated with System-1 (that deals with intuitivity, associativity, and correlation) to improve explainability, generalization and efficiency which is demonstrated through empirical evaluation. This work attempts to improve over existing methods which use limited symbols, and are therefore is insufficient to cover complex patterns of activities for visual activity understanding. They do this by proposing a new system with broad-coverage symbols and rational rules.
Strengths
-This work attempts to solve the lack of compositional generalization (generalized understanding of visual concepts, their relationships, and novel combinations) by creating a system of symbols with broad semantic coverage and rules that are rational and unambiguous. E.g., an existing work contains the following rule: Hand hold ^ Feet close with |= human ride a boat. This rule does not cover the complex patterns of human activities. - This work improves the existing methods by using an inexpensive annotation method by utilizing LLMs (OpenAI GPT3.5 API). This method is able to generate diverse symbols and rules quickly as shown in their symbolic system experiment. For the semantic coverage of the huge image dataset (the target domain), they replace it (the target domain) with the semantic coverage of language models.
Weaknesses
The core message of the paper appears to be that more symbols with greater coverage leads to improved performance. Figure 4 provides a high-level view of how these symbols are generated from LLM output, and Figure 6 notionally shows how they integrate with visual processing, but technical details are missing or unclear. For instance, how are initial symbols "extended into rules satisfying entailment"? In what format are symbols and rules stored? What is the method for integrating the symbolic reasoning system into the overall visual processing pipeline? Their method uses symbol-rule loop to generate new rules whereby they use a known symbol to generate a new symbol and connect it with the rules. The author(s) do not present examples to test the robustness of this method.
Questions
Most questions given above in Weaknesses. From Fig 6, it is not clear if the method will generate the correct set of rules if the images also contained say apples in various circumstances in the background, but the person only decides to buy orange. Can you please clarify this?
Rating
5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.
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
2 fair
Presentation
3 good
Contribution
3 good
Limitations
One limitation regarding symbolic representations of continuous activities is mentioned. The authors have not addressed negative societal biases, and I can think of cases where some societal biases might arise. Activity-based reasoning is at its core a kind of "stereotype"-based reasoning: "Balls are round. Balls roll. Apples are usually round. Apples probably roll." In the domain of common objects, the risks of such stereotype-based reasoning are probably minimal, although if other objects with potentially harmful affordances (e.g., firearms or other weapons) are used without noting that those behaviors are harmful and should be avoided, this presents potential for abuse. Can things like race, gender, or social class be represented symbolically? Should they? Would symbolic representation of these bias system outputs in the presence of "coded" objects like firearms? Beyond universal basic behaviors, who decides what the symbolic representations of them are? Without attending to these questions and more, there is a risk of attributing "typical" or "stereotypical" behaviors to certain individuals or groups of people, and risk generalizing them to other individuals, and hence reproducing patterns of bias and discrimination.
Summary
This paper studies the use of a symbolic system for understanding human activity. The author argues that a good symbolic system for human activity understanding should possess a diverse range of symbols that accurately identify the human activity, and the rules within the symbolic system should be logically correct. The paper proposes the utilization of LLMs to guarantee these two properties and suggests building a comprehensive pipeline incorporating both LLMs and VLMs. Experiments demonstrate that the proposed system outperforms the baseline model, which solely employs VLMs, on multiple datasets.
Strengths
- The paper is well-written and easy to follow. - The paper identifies two ideal properties of a good symbolic system: semantic coverage and logical entailment, which is very reasonable. - Utilizing the commonsense knowledge of LLMs to achieve these two properties is also promising. The experiments have demonstrated the effectiveness of the model on multiple datasets.
Weaknesses
- Some crucial details of the method are still missing: - Regarding symbol initialization, it is not explicitly mentioned how the prompt in Appendix 3.3 (Line 79) is generated. It would be helpful to clarify whether it is manually written or generated using a template, and if so, what is the template. - There are two versions of prompts used for rule extension in the main paper (Line 139) and the appendix (Line 84). In the main paper, a single known symbol is used, while in the appendix, all known symbols are used in the prompt. Additionally, in the appendix, the first sentence contains specific information related to the airplane case: "In a picture, **there is an airplane**". How this particular prompt is generated? - According to Algorithm 1, symbols that contribute to a rule passing the entailment check should be removed from the candidate symbol list (Line 13). However, in Figure 4, after the 1→6→7→8 rule is generated, there are still rules such as 2→1 and 3→1→9 being generated. It would be useful to provide an explanation for this discrepancy. - According to Table 1, some generated symbols are more complex than the target activity, such as "wave goodbye to loved ones," and some do not follow the format of "<The person's hands/arms/hip/legs/feet> <verb> <object>", like "airline staff checking the boarding pass". Given the complexity of the generated symbols, VLMs might face challenges in recognizing them in the images. Although experiments have shown that the current pipeline performs better than directly using VLMs on multiple datasets, there is no guarantee that VLMs can easily recognize the generated symbols. - While the author proposes two ideal properties that should be satisfied, the experiments only focus on the final accuracy of the models on human activity understanding tasks. The accuracy heavily relies on the performance of the VLMs employed. There is no evaluation of how well LLMs ensure these two properties. This makes it unclear what the contribution and bottleneck of the proposed system are. One potential experiment the authors could consider is to have a subset of the dataset and evaluate the pipeline with VLMs replaced by human judgment. But other experiments are also appreciated.
Questions
See my comments in the weaknesses section.
Rating
5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.
Confidence
2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.
Soundness
2 fair
Presentation
3 good
Contribution
3 good
Limitations
The authors have a paragraph discussing the limitations of the proposed method in the last section.
Summary
In this paper, a symbolic-System-2 submodule is integrated with a System-1 module to perform human activity recognition. The work follows the intuitive idea, where a visual prediction module extracts visual symbols and a symbolic reasoning module performs logical reasoning to recognize activities. For the visual symbol extraction module, the work applies a general VLM, to predict the truth value of a predicate in a logic clause. The logical system then performs fuzzy logic to predict the final outcome of the rule, ie, conclusion. The logical system is constructed from LLM, without expensive human annotation in existing works. In experiments, it is shown that the model is better than existing baselines.
Strengths
+ The work leverages and extends the intuitive idea of System-1 and System-2 combination. The integration is important compared to System-1 only pure predictive framework and it's really encouraging to see more and more works integrating these two systems in typical vision tasks. + Compared to existing works, the method leverages LLM for knowledge-base extraction. While the cost saving is from LLM, rather than some nice problem structure design, I do appreciate the efforts to construct formal logic base from natural language systems and make it work.
Weaknesses
I'm not an expert in logical system design so I will refrain from commenting on the system design part on logics. However, I do see some general problems regarding this work. For one, designing logical system specifically for a task risks overfitting. While the performance is improved on a specific task, I do note that the slight increase may not be worth the generality from models like CLIP. Besides, the logic system is pretty complicated. And I do not think the system, or the system construction method could be easily adapted for other tasks. On the other hand, such a system is also only demonstrated in a single task. I doubt if the method could be more general. I'm also interested in seeing more detailed ablation study. In particular, whether it is the System-1 part of the System-2 part that more adversely impacts performance. Would it be possible to assume perfect System-1 and test the System-2 performance or vise versa? Besides, while in general the idea should work. There are many subtleties in telling the truth value of statements, eg, the leg is ``close'' to something. How should such subtleties be addressed? Those being said, I admit the inherent conflict between logical system that are delicate, explainable, robust, and generalizable and deep learning system that are general but fragile. I do not see these points as weaknesses but rather interested in hearing how the authors see them. For my final rating, I'd like to sync with fellow reviewers.
Questions
See 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
3 good
Presentation
3 good
Contribution
2 fair
Limitations
Yes.
Summary
The paper addresses the problem of visual activity understanding and proposes a way to leverage large visual and language models to i) propose symbols from an image, ii) derive new symbols via logic entailment and perform classification. The results show that the augmented system with symbolic reasoning capabilities performs slightly better than the equivalent large models without the reasoning system.
Strengths
The paper proposes an interesting way to leverage large visual and language models to derive a set of symbols relevant to a particular picture and the action being executed in it. Constructing an ontology based on the compiled knowledge represented in an LLM seems to be a powerful approach to enable task specific higher level reasoning.
Weaknesses
One of the key improvements over other related work is claimed to be the broad semantic coverage of the derived symbols and the generated rules. However, given the statistical nature of the used large models it is not clear how accurate these symbols and rules actually are. A deeper analysis of the generated ontology is definitely needed in order to assess the applicability of the proposed method.
Questions
- How is the set of the initial symbols selected? Isn't this initial set generation prone to the same issues as some of the related works? - Given the form of each rule having multiple premises and a single conclusion, why is the hyper-graph structure needed? Wouldn't a simple graph with multiple directed edge be enough? -
Rating
3: Reject: For instance, a paper with technical flaws, weak evaluation, inadequate reproducibility and incompletely addressed 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
2 fair
Presentation
2 fair
Contribution
2 fair
Limitations
The core of the proposed work is based on utilising pre-trained large models which require vast amounts of data and computational resources and so the proposed method is as efficient as described only by assuming access to such large models.
Dear Reviewer K3QA, We are nearing the end of the discussion period with authors. The authors have responded in detail to your review, so pls minimally read and acknowledge their rebuttal, and state which (if any) issues you still do not find to be satisfactorily addressed. You should do so as soon as possible. Thanks, AC
Summary
The paper proposes to increase coverage and generalization of logic-based methods to model 'system-2' or deliberative reasoning for the task of human activity prediction in images. Specifically, the authors incorporate an LLM (GPT3.5) to obtain a larger number of symbols and inference rules related to human activity prediction and further also utilize an LLM to check logical entailment of generated rules (based on which noisy rules are filtered). Their proposed symbolic system operates upon a VLM which is used for symbol extraction from the image, based on which rules are activated with a fuzzy-logic computation (to obtain a probability score for the inference chain). Authors show that adding their method on top of two VLMS (CLIP and BLIP2), enhances their performance on prominent activity recognition benchmarks and also show visualizations of activated rules and predictions besides ablation studies of their method.
Strengths
1. The usage of LLMs to broaden semantic coverage of symbolic/logic-based methods by discovering relevant premises and rules and also for checking/approximating logical entailment is relatively novel. 2. The methodology and formalism of their approach is well described, with appropriate examples and figures/tables to help reader understand the approach. 3. Results on multiple prominent benchmarks (including HAKE, HICO and Stanford-40) for human activity prediction show that the method can benefit both zero-shot (for BLIP2 and CLIP) and finetuned (for CLIP) performances of prominent VLMs. Ablation results also highlight contributions of individual method components/variations.
Weaknesses
1. The symbolic system's application draws on using a VLM multiple times to extract the probability of individual symbols. It thus seems that the computation will linearly scale with number of symbols, which can be a significant drawback due to the incurred computation cost. (VLMs such as BLIP2 by themselves have very large computation cost, hence even increasing that 100x for 100 symbols is a very significant additional compute). Further, extending this to further domains beyond human-activity recognition would require many more symbols, and ultimately possibly not be scalable. 2. Evaluation baselines only consider results of the backbone VLMs currently. There are multiple recent works in related direction of human-object interaction and activity prediction in zero- or few-shot scenarios which have not been mentioned or used as baselines (e.g. Weakly-Supervised HOI detection (ICLR 2023); RLIP (NeurIPS 2022)). Currently only RelVIT is mentioned in supplemental. Authors should either compare with such methods or mention why these recent methods are not used/suitable for comparison. 3. Further, since LLMs are employed for much of the 'system-2' reasoning, authors should consider reporting results of alternative methods using VLMs + LLMs directly without any logical system (e.g. VLM generates image caption, based on which LLM predicts possible activity when prompted for each activity in the label space -- similar to PICa in "An Empirical Study of GPT-3 for Few-Shot Knowledge-Based VQA" AAAI 2022). These results could better highlight the contribution of the proposed logical system and that performance increments are not due to just LLM-VLM combination. Less important: 1. Failure cases could be illustrated for a small sample of the questions to better highlight which component is currently leading to errors (e.g. is it the VLM not producing accurate symbols or is it a noisy rule, etc)
Questions
Please see weaknesses. My primary concerns (impacting rating) are regarding: 1. Does compute scale linearly with number of symbols and would this be a major limitation in practice? If so, can this can be addressed through alternative symbol grounding/identification methods? 2. The relatively limited evaluation compared to other relevant works for human activity prediction. How does the method compare to related baselines in human-object interaction learning such as ICompass (Interaction Compass: Multi-Label Zero-Shot Learning of Human-Object Interactions via Spatial Relations (ICCV 2021) and RLIP (NeurIPS 2022)) or alternatively multimodal LLM methods (e.g. PICa in "An Empirical Study of GPT-3 for Few-Shot Knowledge-Based VQA" AAAI 2022)?
Rating
5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.
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
2 fair
Limitations
Potential negative societal impact is mentioned to be minimal.
Summary
The paper proposes a novel method for visual activity understanding using an enriched symbolic system. The authors focus on enhancing the integration of intuitive (System-1) and logical (System-2) reasoning processes in AI systems. They argue that the current System-2 reasoning methods suffer from limitations such as a dearth of symbols and limited, ambiguous rules, inhibiting the system's ability to capture the complex patterns of activities and limiting its explainability, generalization, and data efficiency. In response to these shortcomings, the authors introduce a new symbolic system enriched with a broader set of symbols and more rational rules. This system, underpinned by commonsense knowledge, aims to foster a more comprehensive understanding of human activities visually. To support their work, they leverage the advancements in LLMs and techniques like the symbol-rule loop and entailment check.
Strengths
The concept of developing a new symbolic system that combines aspects of intuitive (System-1) and logical (System-2) reasoning is indeed a creative approach. However, the idea of incorporating commonsense knowledge into AI reasoning is not entirely novel, as there have been previous works exploring similar notions. The authors put forward a detailed methodology and apply recent advancements such as Large Language Models (LLMs) in their approach. While the paper is mostly well-written, certain parts, especially the description of the symbolic system and its instantiation, could benefit from more elaboration or clearer explanations. Although the paper targets a significant area in AI - the integration of System-2 reasoning into visual activity understanding - it's unclear how much practical impact this specific method would have on the broader field. The lack of comparison with other methods, especially in various real-world applications, makes it difficult to evaluate the actual significance of their work.
Weaknesses
The paper's evaluation approach, while it encompasses multiple datasets, does fall short in providing a comprehensive set of comparisons with current state-of-the-art methods. The selected baselines do not reflect the full breadth of existing approaches, which undermines the claim of superiority of the proposed method. It would be valuable to include comparisons with more recent vision-language models, such as LLaVA, MiniGPT-4, and LLaMA Adapter, to provide a more meaningful assessment of the proposed system's performance. The paper does spend a substantial portion of its content introducing the method background. Shifting sections 3.1-3.2 to the appendix could help maintain focus on the main contribution of the paper, while still providing the necessary context for interested readers. The explanation of the core methodology is another area that requires improvement. It is not adequately detailed in the introduction, with key aspects of the proposed method encapsulated in a single sentence (lines 57-58). The paper would benefit from a more thorough and step-by-step explanation of the modeling method in the early part of the paper, such as the introduction, which would help readers better understand the unique features and operations of the proposed system.
Questions
Missing related references: - Knowledge Aware Semantic Concept Expansion for Image-Text Matching, IJCAI 2019. It would be beneficial to include more latest baselines for comparison.
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
3 good
Presentation
3 good
Contribution
2 fair
Limitations
The authors adequately addressed the limitations.
Dear Reviewer sNaH, We are nearing the end of the discussion period with authors. The authors have responded in detail to your review, so pls minimally read and acknowledge their rebuttal, and state which (if any) issues you still do not find to be satisfactorily addressed. You should do so as soon as possible. Thanks, AC
Dear Reviewers, The authors have responded in detail to your review feedback, and the current reviewer-author discussion phase ends on Aug 21. If you have not already done so, pls read the rebuttal(s) and respond within the **next 2-3 days**, so as to allow for further discussion (if needed). You should read the authors' rebuttal(s) in detail, and ideally: 1) Acknowledge the points that you find have been satisfactorily addressed; 2) Ask for further clarifications where needed; 3) Explain why you may still find certain issues to be insufficiently addressed Thanks, AC
Reply to the rebuttal
I thank the authors for addressing the concerns adequately. Regarding the motivation of probabilistic reasoning, it would be beneficial to add the literature on probabilistic (logic) programming [1,2] since they essentially have the same motivation as explained in the rebuttal, i.e. dealing with uncertainties on symbolic reasoning. Describing the difference/novelty compared to them would clarify the contribution of the paper more. [1] Luc De Raedt, Kristian Kersting, Sriraam Natarajan, David Poole: Statistical Relational Artificial Intelligence: Logic, Probability, and Computation. Synthesis Lectures on Artificial Intelligence and Machine Learning, Morgan & Claypool Publishers 2016 [2] Daan Fierens, Guy Van den Broeck, Joris Renkens, Dimitar Sht. Shterionov, Bernd Gutmann, Ingo Thon, Gerda Janssens, Luc De Raedt: Inference and learning in probabilistic logic programs using weighted Boolean formulas. Theory Pract. Log. Program. 15(3): 358-401 (2015) As the concerns are addressed, I raise the score to 6 (from 5).
Thank you for your feedback!
We thank the reviewer for the feedback! We will add the literature on probabilistic (logic) programming [1,2] and describe the difference/novelty compared to them: The standard probabilistic (logic) programming typically solves simplified relational tasks, e.g., the WebKB dataset, and the BoxWorld problem. In comparison, we focus on a real-world vision task that involves complex patterns (rules). Thus, a novel symbolic system has been proposed.
Thanks for the clarification. The response is satisfactory. I will keep my accept vote unless other reviewers point out critical issues I missed.
Thanks for the clarification and the additional experiment results. From the response, the prompts for symbol initialization and rule extension are manually designed for different tasks and datasets (e.g. the symbols are constrained to hand-related states for HOI recognition). I have a number of questions regarding these prompts: 1. What are the prompts for every dataset/task? It is very important to include the original prompts in the paper for the purpose of reproducibility and clarity. 2. How are these prompts determined? 3. Are experiment results sensitive to the design choice of prompts? I am a little concerned about how much extra effort we need to tune the prompts to have a good performance. I understand that it is hard to have a systematic way of evaluating the effect of prompts. And it might be hard to generate prompts that generalize across different domains. But it is valuable to include discussions and details for a paper that heavily leverages LLMs, because LLMs are known to be sensitive to prompt design.
We thank the reviewer for the feedback. Questions are answered below. ### What are the prompts for every dataset/task? The prompts for the proposed dataset/task related to activity understanding are basically similar. The prompt differences are small modifications necessary to adapt different dataset/task settings: - When an interacted object is known, it is listed in the rule extension prompt for emphasis. - For HAKE-PaSta, the symbol initialization is omitted. (suppl L147-150) Also, we respectfully point out it is a misunderstanding that "hand-related states" is "tailored for HOI", according to our response above: "...constrain the answer symbols as hand-related states since they are more common than other symbols...". We attach the original prompt in Appendix 1 in the following response for further investigation. ### How are these prompts determined? The prompts are determined via a trial-and-error process by human experts on randomly sampled 10 activities. Human experts have read the image samples of these activities, and thus have expert knowledge about what symbols/rules are more suitable for visual reasoning. Each round in the trial-and-error process is as follows: - one expert writes a prompt - an LLM generates answers by the prompt - all experts judge whether the answer is suitable - all experts modify the prompt based on the discovered problems An example of the trial-and-error process is detailed in Appendix 2 in the following response. We admit the current process needs improvement: the human experts are limited to the authors, and the prompts are verified on a few activities. However, it is a performance-cost trade-off: as the reviewer points out, it takes extra effort to tune the prompts to have a good performance. ### Are experiment results sensitive to the design choice of prompts? Experiment results are sensitive to prompts with major differences. Using the older version prompt during the trial-and-error process brings difficulty to correct and automated symbols generation. Experiment results are less sensitive to prompts with minor differences, i.e., slightly paraphrasing the prompts. We conduct a brief experiment: - get another 2 paraphrased symbol init prompts and another 2 paraphrased rule extension prompts - use these $3*3=9$ prompts to generate 9 sub-systems - measure the average pairwise graph similarity - the average similarity (range: [-1,1]) on 10 randomly sampled activities is 0.83
### Appendix1 For an activity with an interacted object (e.g., board an airplane) in HICO, HAKE-Verb, Stanford40: - (Symbol Initialization) In a picture, a person is boarding an airplane. What are his hands doing? Answer with 5 concise, highly-related phrases. The format is "<hands> <verb> <object>". - (Rule Extension) In a picture, there is an airplane. IF [The person is holding a boarding pass...] AND [condition] THEN [The person is boarding the airplane.] [condition] is one concise phrase. The format is "<The person’s hands/arms/hip/legs/feet> <verb> <object>". What is [condition]? For an activity with no interacted object (e.g., cook) in HICO, HAKE-Verb, Stanford40: - (Symbol Initialization) In a picture, a person is cooking. What are his hands doing? Answer with 5 concise, highly-related phrases. The format is "<hands> <verb> <object>". - (Rule Extension) In a picture, IF [The person is holding a slice...] AND [condition] THEN [The person is cooking.] [condition] is one concise phrase. The format is "<The person’s hands/arms/hip/legs/feet> <verb> <object>". What is [condition]? For a PaSta (e.g., hold something in both hands) with a known activity (e.g., wash an airplane) in HAKE-PaSta: - (Symbol Initialization) None - (Rule Extension) In a picture, IF [The person is washing an airplane...] AND [condition] THEN [The person is holding something in both hands.] [condition] is one concise phrase. The format is "<The person’s hands/arms/hip/legs/feet> <verb> <object>". What is [condition]? ### Appendix2 Trial-and-error process of the prompts for symbol initialization Round1: - Prompt: In a picture, a person is boarding an airplane. What are useful visual features to distinguish his activity? - Answer: Useful visual features...could include: 1. Luggage: Look for the presence of suitcases, backpacks, or other types of bags that suggest the person is preparing to travel... - Comment: Object-centric, difficult to extract key information about human activities. Round2: - Prompt: In a picture, a person is boarding an airplane. What is he doing? - Answer: The person in the picture is boarding an airplane. Boarding refers to the process of entering the aircraft before the flight begins. It typically involves presenting a valid boarding pass, showing identification, and walking through the designated gate or bridge to reach the plane. - Comment: Verbose. Round3: - Prompt: In a picture, a person is boarding an airplane. What is he doing? Answer with 5 concise, highly-related pharses. The format is "<human> <verb> <object>". - Answer: Person boarding airplane. - Comment: The answer repeat the question. Round4: - Prompt: In a picture, a person is boarding an airplane. What are his hands doing? Answer with 5 concise, highly-related pharses. The format is "<hands> <verb> <object>". - Comment: Satisfactory.
Thanks authors for the response. I think the current system provides a reasonable way to leverage LLMs and VLMs to do visual activity understanding. But according to the experiment results, the prompting method can critically affect the performance, limiting the robustness and generality of the system. I will keep my current rating.
We thank the reviewer for the follow-up feedback. The suggestion provides a deeper analysis of our method. Nevertheless, we hold a slightly different view about the prompting method: - Prompting affect the performance: We leverage a trial-and-error process to find the proper prompt. Minor modification on the final prompt has a relatively low impact on the topology of the symbolic system. - Robustness: Modified prompts lead to different results. This is not necessarily a limitation of robustness but can be a way to increase diversity, i.e., various reasonable symbols/rules can be generated via prompting without much extra effort. - Generality: The trial-and-error process can be reused by other tasks. We mainly focus on the insight of two properties in our initial submission and did not preferentially consider the prompt design, which is an alternative implementation. We will add more analysis and discussion about the impact of prompting. Thanks again for your suggestion!
Thanks for the detailed response and clarifications. Regarding point 1 on computation scalability -- I agree that a hierarchical structure as described should reduce computational cost. However, it is unclear whether this is currently integrated in the proposed framework and reported results? Or is it a proposed extension for scalability? In relation, a quantitative comparison (e.g. number of operations/FLOPs) could be beneficial and provide more clarity on this potential issue. It is also unclear to me whether the probabilities of symbols are to be retrieved for each potential activity class? E.g in Fig. 6 Image I-IV, currently relevant symbols are shown for 'buy orange' class; would the probability retrieval also be done for other relevant symbols of other possible classes (such as 'visit store', etc) for each image? (in this case, it seems the number of total symbols would be larger than 5 and scale with number of classes) This is my remaining major concern. My other concerns have been addressed adequately. The VLMs + LLMs ablation results could be added in the main paper to highlight the symbolic system's benefits. Similarly, quantitative analysis on System-1 errors and System-2 defects could be added to main paper.
We thank the reviewer for the feedback. We will update additional results and analysis. Questions are answered below. ### The hierarchical structure The hierarchical structure is a proposed extension for scalability. We mainly focus on the effectiveness and implementation of our insight in our initial submission and did not preferentially consider the efficiency. Nevertheless, it is easily integrated into the proposed framework, with a new hierarchical connection applied to the original symbolic system. ### Are symbols retrieved for each potential activity class? The total number of symbols to measure will increase with the number of potential activity classes, but not scale linearly. - It is unavoidable to measure probabilities of more than one activity class for activity understanding tasks. This is because activity recognition is a multi-label classification rather than a multi-class classification: A person typically performs multi-action simultaneously, e.g., standing while eating. Thus, the increase in symbol measurement comes from the inherent characteristic of activity tasks. - However, the symbols are reused for different activity classes without repetitive computation. For example, "buy orange" and "visit store" shares the symbols: ($m_a$) Interact with a seller, ($m_d$) Payment process, ($m_1$) Talk with seller, ($m_{10}$) Give money to seller, ($m_8$) Reach for a wallet. It is also illustrated in the main paper Fig.2(a): the red and green activities share some symbols (gray), e.g., second from right, third from left. ### Quantitative comparison To improve clarity, we report the num-of-operations average per image on HAKE-Verb. An "operation" is defined as predicting a symbol or an activity class directly. | reasoning | num-of-operations | | ----------- | ----------- | | w/o | 18 | | w/ | 39 | Without reasoning, activity classes are predicted directly, and num-of-operations vary from images due to different ground-truth object labels which verb prediction conditions upon (suppl L138). With reasoning, reused symbols are predicted via a hierarchical structure. We provide a more detailed analysis of a small case. There are 38 images known to contain oranges, where verbs (buy/cut/eat/hold/inspect/peel/pick/squeeze/wash) are to be predicted. Without reasoning, num-of-operations=9 (same for each image). With reasoning: | reuse | hierarchical | num-of-operations | remark | | ----------- | ----------- |----------- |----------- | | w/o | w/o | 71 | same for each image | | w/ | w/o | 31 | same for each image| | w/ | w/ | 23 | average over 38 images|
Thanks for the further details and clarifications. Based on the number of operations, I believe the additional computation is relatively significant (as mentioned 18 operations w/o reasoning to 39 with reasoning, which is more than double). Further, when applying method to new domains, additional symbols and rules would need to be extracted/identified which is dependent on underlying VLMs and LLMs and is more complex than direct finetuning / few-shot application of VLMs. While these may be addressable through future work, I encourage authors to explicitly highlight these points on computation scalability and applicability to other domains (with dependence on VLMs/LLMs) in the limitations section of their current work. I have raised my score to 5 given the work's novelty and performance benefits.
We thank the reviewer for the insightful suggestion! We will discuss computation scalability and applicability to other domains in the limitation section, and explore these important issues in future work.
Thanks to the authors for their response. Most of my questions are addressed. I have a follow up regarding the rule in Figure 6: what you are doing here is exchanging one bias for another, namely exchanging the bias toward detection of background objects as would be present in a pure vision system with a bias toward the stereotypical reasoning I mentioned in the Limitations discussion (that is, the rationality of the rule will bias the output). In the example you provide, this seems to work well (and I would recommend that you add your answer #4 to the paper for clarification purposes), but there may be cases where this does not improve the results or produces incorrect output (e.g., the presence of a bicycle might bias the output toward "human riding bicycle" even if the person is just standing next to it—see the discussion of false positives in Gkioxari et al "Detecting and recognizing human-object interactions"). This is okay but I would recommend that you be clear about this. System 1 vs. System 2 reasoning is a cost-benefit tradeoff and simply integrating the two is no guarantee that you nail the right combination for globally better performance. I will keep my score.
Thanks for pointing out the potential bias brought by symbolic reasoning. The suggestion is inspiring. We will investigate deeper about the difference between System 1 vs. System 2 performance and better integration, which is an interesting and under-explored problem. We will add more discussion and clarification about Fig.4 visualization results.
Dear Authors, On behalf of the reviewers who have not responded (yet), thank you for the detailed responses. After reading the reviews and rebuttals, I find the clarifications (about symbol selection, recent VLMs, etc.) and further analyses (e.g. bottleneck analysis) to be fairly reasonable. I note the additional results for CIFAR100 and VCR, along with results for Otter. I will give all of these proper consideration. I do have one general and fairly minor query/comment, which may not be specific to your work alone (but symbolic methods in general). Is it the case that the nature of symbols forces a certain commitment to a fixed, discrete concept (e.g. "hold"), orthogonal to the issues of probabilities and fuzzy logic? More specifically, how can subtleties in the image and/or language modality be handled properly? Using "hand hold onto the side of the boat" as an example, how can/do symbolic methods handle subtleties and variations like "hand hold tightly" vs. "hand hold gently" vs. "hand resting on", which may affect downstream reasoning outcomes? Thanks, AC
Dear AC, Thanks for your valuable time and efforts and your encouraging feedback. We answer the questions below. The nature of symbols: - Traditional symbolic methods treat one symbol as a fixed, discrete concept in simplified relational tasks (e.g., the BoxWorld problem) where symbols are definitely True/False, e.g., "box1 on box2". - For deep-learning-based visual recognition tasks, symbols are more complex. - About subtleties: Visual concepts (symbols) are hierarchical. Based on one symbol (e.g., "hand hold"), we can define some finer-grained symbols (e.g., "hand hold tightly") belonging to it. - About variations: Symbols have different semantic correlations. For example, "hand hold" is more related to "hand rest on" than "hip seated". - About probabilities and fuzzy logic: The symbols are not definite because there is no pre-defined world model. Thus, the model predicts the symbol probabilities and derives the conclusion via fuzzy logic. Fuzzy logic can be replaced with a gate function applied to probabilities, but it will bring a performance drop due to information loss. How to handle subtleties and variations in our paper: - For the symbolic system construction with an LLM, we use semantic embedding distance as a threshold to judge whether two text descriptions are treated as the same symbol (Alg. 1, L13: redundant symbols removed). Under this standard, "hand hold tightly" and "hand hold gently" are the same symbol as "hand hold", while "hand rest on" and "hand hold" are treated as two separate symbols. In most cases, it avoids activity confusion (detailed in our 1-st response for reviewer yZAX, "Contribution\&Bottleneck Evaluation".2). - For the symbol probabilities of a specific image, the prediction depends on the ability of the adopted System-1-like VLM, e.g., finer-grained verb recognition, and adverb recognition. Possibility of finer-grained symbols: - Better performance may be achieved by a symbolic system with finer-grained symbols, as well as more powerful VLMs to recognize finer-grained symbols. - But it is a performance-cost trade-off. A symbolic system with finer-grained symbols will increase computation costs. Improving a VLM's ability for finer-grained recognition needs more densely annotated data (e.g., adverb dataset [1], visual captioning or VQA datasets, pseudo labels from VLMs, etc.). - We regard it as interesting and important future work to analyze and leverage finer-grained symbols to facilitate reasoning. Thanks again for the insightful feedback! [1] Pang, Bo, et al. "Further understanding videos through adverbs: A new video task." Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 34. No. 07. 2020. Thanks, Authors
Decision
Accept (poster)