What's Left? Concept Grounding with Logic-Enhanced Foundation Models

Summary

This paper introduces Neuro-FOL, a method combining a large language model for generating FOL programs, an FOL executor and trainable concept grounding modules to improve performance on a number of visual/3D QA style tasks. In the method, a LLM interpretor is used to generate a FOL program that is not specific or grounded to the exact domain. A differentiable first-order logic executor is then used to hierarchically evaluate the value of the program by using learnable, domain-specific grounding modules to ground the program in the specific domain. The paper evaluates in 4 settings: 2D question answering, 3D referring expressions, temporal sequence reasoning and robotic manipulation. The method performs similarly to both NS and end-to-end methods in the normal data settings shown in Table 2, beats end-to-end methods in low data settings (Table 4). The paper also shows good performance on a set of held-out CLEVR tasks.

Strengths

I think main idea of the paper is really clever and well motivated. I think using a pre-trained LLM to do the logical breakdown of the query by generating a program, which can then be differentially trained is really smart. It gives you a way to have both the generality you get from LLMs having seen many concepts and the flexibility of adapting to a new domain. The wide variety of domains and datasets evaluated is really good. Shows that the method works across a wide variety of different domains. Shows really good performance compared to end-to-end methods in the rare data case, which is good. Paper is pretty clearly presented and motivated.

Weaknesses

I think the claim of it being a universal language for all domains and therefore generally applicable across domains is maybe misleading and not totally supported. I think I agree that you could generate a program for pretty much any query-based problem, but it sort of missing a critical assumption, that the method has sufficient data to train each of the domain-specific module that you would generate. And of course, if during evaluation, if the program generated a module that it hadn't seen during training, it would fail. I think this is a real limitation of the method that should be more clearly mentioned in the paper. Related to the above point, the set of domains is somewhat limited, and especially the kinds of reasoning here are pretty narrow. For instance, the 2D question answering is all done on CLEVR where the visually domain is extremely limited (only a set number of different block objects in a set number of colors) and a rather limited space of kinds of questions (questions chaining together reasoning about objects color, relative position, count etc). There are for instance many kinds of questions about images (for instance in the VQA dataset) that might not really have good coverage in the training set to work. If you had to train a new MLP classifier for each kind of attribute, object class and relation you might find in VQA, it's not totally obvious to me that this method would have enough data to actually work. SR3D is definitely more visually diverse in the number of object types, but the question types are also quite limited to these kinds of spatial reasoning questions. Similarly, robotic manipulation is more visually diverse and in the number of objects, but also suffers from a lack of diversity in the kinds of queries it needs to perform. I get that the advantage of this method over pure NS approaches is that you do not need to manually define a domain-specific language, but the comparisons don't really any improvement over them. The claims about universality would be stronger if the domains were less constrained, showing for instance that this method can deal with unusual domain groundings in a way that would be hard to anticipate for a designer for NS methods. Kind of a minor point, but I think the point on line 105 about VLMs not working on different domains such as 3D scenes and motion sequences is pedantic. Models not trained on 3D scenes and motion sequences of course don't generalize to these, but one could train a network on these domains as well. For instance: Gato (S Reed 2022) is trained on multiple kinds of modalities. "Flamingo" in Table 6 should REALLY be stated to be Open-Flamingo. This is especially confusing since the paper uses the citation when it names the method. But it's not using the Flamingo model, it's using an open-source alternative to the model not Flamingo itself, which is very confusing/misleading. A couple of unanswered questions I had (see below).

Questions

In Table 6, is Neuro-FOL trained on the regular CLEVR task before being put in the held-out CLEVR tasks? What about [Open] Flamingo? In general, could you more clearly state in the paper the training loop for Neuro-FOL. I sort of infer that it is trained end-to-end on the target tasks, but that the only parts of the network that are actually updated are the domain specific grounding modules? Is that correct? How does your method respond to very rate concepts, if you need to finetune an MLP for every grounding function. Do you see cases where a concept has very few examples during training and that module performs poorly? How often does the LLM fail to provide a valid program and how often does it fail to provide a correct program? Does an invalid program result in Neuro-FOL failing to run? If the program is incorrect during training, does this hurt performance in eval? How robust is the method in general to program failures? In general, what are the failure cases of this method?

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

2 fair

Presentation

3 good

Contribution

4 excellent

Limitations

I don't think that the paper does much to address the limitations of their methodology (see above) which I think would add a lot of value to the paper.

Summary

The paper presents an approach for solving visual reasoning tasks across multiple domains such as 2D image QA and robotic object manipulation. The approach (Neuro-FOL) uses an LLM to generate a first-order logic (FOL) program that is executed with a combination of learned and predefined modules. One aim is to take advantage of a pre-trained LLM (e.g., GPT 3.5) to generalize to new visual reasoning tasks at inference time. Neuro-FOL is evaluated on a diverse set of benchmarks including the CLEVR 2D VQA tasks, ReferIt3D 3D spatial reasoning tasks, HumanMotionQA tasks, and the CLIPort object packing tasks. Results demonstrate that Neuro-FOL is competitive with and sometimes outperforms strong baselines.

Strengths

- The problem of grounding concepts across different domains (e.g., 2D images vs. 3D scenes) is an important problem because many concepts are naturally defined across these domains. This problem is well motivated in the introduction. - The ability of Neuro-FOL to learn new concepts that are generated by the LLM during training is neat. Practically speaking, this should save designers the time required to define these concepts. - The paper evaluates Neuro-FOL on a diverse set of domains including 2D image QA, 3D scene QA, motion capture QA, and robotic manipulation, which demonstrates the generality of the proposed approach. Furthermore, Neuro-FOL has strong performance in all tasks, suggesting that it will similarly do well in other domains. - The paper is well written and easy to follow, but some details are missing as discussed below.

Weaknesses

- The abstract and introduction start with the grand vision of grounding the concept “left” in multiple domains. However, from the experimental evaluations, it is unclear if such “domain-independent” reasoning is (a) learned and (b) useful. - (a) Are concepts learned in a domain-independent manner? In other words, is Neuro-FOL jointly trained on datasets from multiple domains? These details are not clearly stated in the current manuscript. - (b) The experimental results do not clearly demonstrate if “domain-independent” grounding is useful. A specific experiment to test this would be training Neuro-FOL independently on different domains and then comparing the performance of Neuro-FOL training jointly. Such an experiment does not appear in the paper. Thus, I do not see support of these claims. - The proposed approach differs from LLM + API based methods (e.g., Gupta and Kembhavi, 2023) in two ways: (a) the use of first-order logic and (b) not requiring a predefined set of modules (i.e., concept networks in Neuro-FOL). It is unclear which of these two design concepts is most important. We can imagine a world in which (a) does not matter but (b) is critical. Alternatively, both (a) and (b) might be important. The experimental evaluations do not disentangle these two features. Thus, it is unclear where the important novelties in the method lies. Note: the ViperGPT experiments do not disentangle this question. Thus, it would be helpful if the authors could comment on this further. - While learning concept MLPs through Neuo-FOL training has the benefit that concepts do not need to be predefined, this may also be a disadvantage because there may only be a small amount of data for a given concept. For example, a given dataset may have many questions asking about “left” and “right” and may only have a few asking about “in between” or “sandwiched between”. It is unclear how the proposed approach would generalize to such situations. - L257: “Notably, Neuro-FOL is able perform on all domains, instead of requiring specific models for each setting.” This claim is misleading, as components of Neuro-FOL are trained for specific domains.

Questions

1. How does Neuro-FOL performance compare when training jointly on multiple domains vs. independently on individual domains? 2. Are both (a) using FOL and (b) not having a predefined set of concepts important? 3. How does Neuro-FOL perform on concepts that are not well represented in the training data?

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

2 fair

Presentation

3 good

Contribution

3 good

Limitations

The limitations are briefly discussed, but additional details could be provided (perhaps based on answers to the questions above). The broader impacts are discussed in the appendix.

Summary

This paper proposes an approach for general-purpose language grounding across a variety of domains and tasks. The approach first uses an LLM to generate a domain-general program, which can be implemented across different domains using domain-specific functions, represented as neural networks (and domain-general implementations of certain FOL predicates). Experiments are conducted on a variety of domains, including image understanding, video understanding, and robotic manipulation.

Strengths

The approach is compelling and relatively original -- it is especially original in combining a single unified formal representation across a variety of tasks. It is obviously addressing an important problem, and having modular approaches like these is important for investigating failure cases. The approach is quite clearly detailed, except for details in training. Experiments are very thorough. I appreciate evaluating on human data. I would have liked to see more analysis in model performance.

Weaknesses

My main concern is that modularizing the approach into essentially two components: (a) language -> FOL, and (b) FOL + perception -> answer/action, is glossing over nuances of language use that may require interaction between the language and perception aspects. For example, the implementation of "left" may be dependent on the identities of the objects participating in the spatial relationship: there may be a canonical "left" side of an object, given its affordances (e.g., the "left" of a fridge may be the area to its left when one is standing facing the fridge, regardless of the actual perspective being taken in the 3D question answering task); it seems the current approach would not be able to handle this context-dependence of "left", unless the domain-specific implementation of "left" was somehow able to encode all of the information necessary to make this inference (that a particular region is a fridge, and that a fridge has a canonical "left"). Another place where this comes up is that the meaning of adjectives can be highly context dependent in natural language use: e.g., the meanings of color terms depend strongly on context (as a specific example: let's say there is a very pale blue object and a very pale red object -- when hearing "blue", one would most likely choose the pale blue object. But if the same very pale blue object is paired with a more vibrant blue object, "blue" is most likely going to refer to the more vibrant one). In general, I'd be interested to hear discussion about this particular problem and how (if) the proposed approach could implicitly capture these more direct relationships between language use and perception, or if FOL acts as a bottleneck which removes any ambiguity in sentence meaning (it seems this is the goal in its design -- even the program \iota x blue(x) ^ sphere(x) in L132 may be ambiguous if there are multiple blue objects of varying shades!). I also would have liked to see more in-depth analysis of the approach. In particular, what's left for tasks where there is still a lot of room for improvement -- basically everything except CLEVR? Because the approach is modular, it seems there is a lot of room for analysis in how different parts of the approach might be failing -- are the programs incorrect? (What causes a program to be inexecutable?) Are the learned domain-specific functions erroneous? Small nitpicks: * Having a consistent example in Figure 2 would be nice. * I'd suggest putting the FOL programs in Figure 3

Questions

* Is using MLP implementations of domain-specific functions limiting in some ways? I.e., are there functions which, in experiments, are more difficult to train than others? * Are the implementations of the FOL primitives hardcoded (e.g. as "min" for "and")? * How are the domain-specific functions trained? There are few details on the training process, but I am assuming as it is end-to-end differentiable that they are trained using LLM-generated programs on some training data * How is the arity determined for concepts? Is it just "proposed" by the LLM during training? * I was confused about the presentation of Table 4 and 5 -- what is a data efficiency setting? Is this just the amount of training data used (presumably to train the domain-specific functions)?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

There wasn't much discussion of limitations. I would have liked to see discussion wrt. analysis of the proposed approach, and the potential limitation of completely separating the language and perception aspects of these tasks by a single FOL program.

Summary

The manuscript proposes a framework that leverages an LLM to map queries to executable first-order-logic programs, with domain-specific grounding functions. The manuscript includes experiments on multiple tasks and domains.

Strengths

The paper is mostly well-written. The paper considers an interesting topic, under compelling methodology (i.e., neuro-symbolism). The presentations of visualisations and results are mostly clear.

Weaknesses

Section 1 (Introduction) / Figure 1 — One has to be particularly careful in applications that involve some (other) embodiment, as concepts like “left” could adopt an alternative reference frame, beside just the egocentric one. How does the approach incorporate this reasoning? Section 1 (Introduction; L49-51) — Whereas no domain-specific language is required at the reasoning level, there still needs to be lexical alignment with the concepts supported by the domain-specific grounding modules. Furthermore, the approach does still need predefined built-in functions (Table 1) and domain-specific concept names (L158). Section 3.2 (L175-177) — Let’s be exceedingly explicit here. The manuscript states that the function implementations do not require manual definition in code. How are they generated? What are all the core FOL programs? What are their mappings from concepts initialized by the LLM? Table 4 / Table 5 — The data percentages seem to be chosen arbitrarily. How were the percentages decided on? What happened for, e.g., 25% and 50%?

Questions

N/A — 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

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

The manuscript does not include any sections on Limitations or Societal Impact.

Official comment by Reviewer 2ce2

I appreciate the authors' detailed responses to all reviews. I have no further comments, and I am satisfied with the answers to my questions; I will increase my score by one point.

Happy to answer any further questions

Dear Reviewer B8BR, Thank you for reviewing our submission. We have posted our response per your suggestions and questions. We are happy to discuss with you and answer any further questions. As the deadline for discussion is approaching, we very much look forward to your feedback. Thank you, Authors

Happy to answer any further questions

Dear Reviewer DugZ, Thank you for reviewing our submission. We have posted our response per your suggestions and questions. We are happy to discuss with you and answer any further questions. As the deadline for discussion is approaching, we very much look forward to your feedback. Thank you, Authors

Happy to answer any further questions

Dear Review xPiC, Thank you for reviewing our submission. We have posted our response per your suggestions and questions. We are happy to discuss with you and answer any further questions. As the deadline for discussion is approaching, we very much look forward to your feedback. Thank you, Authors

Response

Sorry for the late reply. I think the response answered a lot of my questions. Given that and that I liked the paper to begin with, I will increase my score by one.