By Anonymous User
Review Details
Reviewer has chosen to be Anonymous
Overall Impression: Weak
Content:
Technical Quality of the paper: Weak
Originality of the paper: Yes, but limited
Adequacy of the bibliography: No
Presentation:
Adequacy of the abstract: No
Introduction: background and motivation: Bad
Organization of the paper: Needs improvement
Level of English: Unsatisfactory
Overall presentation: Weak
Detailed Comments:
This paper proposes a formal method for relating symbolic data structures to numeric representations or embedding them in metric space by using a parameterized edit distance, enabling the definition of a geodesic between data instances. The submission further proposes the definition of a data region in metric space. Apart from the formal specification, the authors propose to illustrate the practical implementation of their specification in toy demonstrations as well as a full experiment.
The proposed idea is relevant to the NeSy community, as one important step of combining symbolic representations with neural representations is to obtain numeric representations of the former. However, what is missing from this paper is a convicing idea or even a proper discussion on how to connect the proposed numeric representation of symbolic knowledge with neural representations. Instead, toy examples of replacing other representations with the proposed one and directly performing edit distance-based transitions of objects, e.g. images or audio, and a combination of this representation with a statistical reinforcement learning approach are provided. It is left open, how this might be effective for a neuro-symbolic approach. The problem solving indicated in the abstract and title seems to be attributed a more than secondary role in the article itself, only illustrated in a toy example towards the end. Given these issues of objective paired with the substantial scientific standards and readability issues detailed below, I believe that this submission, while for sure having its merits, is not suitable for this journal.
The abstract and introduction need to be strengthened and more closely related to the objectives and contents of the proposed paper. For instance, the definition of a symbol in the introduction very quickly goes from one idea and definition to the other, relying largely on the references for a proper context of the discussed ideas. For instance, how does symbolic grounding yield the notion of consciousness and how is this discussed in the reference of Logic Tensor Networks provided for this claim? Why is there an invisible Wikipedia link on the word affordance? I believe links should first be visible and second to scientific sources.
The argumentation on why neural methods might benefit goes from relating symbolic representations to numeric representations straight to this "occurs for explainable AI". It then even goes into claiming that such a relation might be an effective "way to model cognitive representations", which seems like an overstatement. I believe that the argumentation in the abstract and introduction should be clarified, cleaned, and strengthened. Both should also present the main contributions of the submission and work.
The clarity of the neurosymbolic approaches section would equally benefit from a very thorough revision. The section goes from talking about machine learning on the level of Bayesian methods over a claim on how important symbolic approaches are for explainability to the ecological footprint of "deep network computation", meticulously avoiding the word neural or neural network. Even in describing the taxonomy presented by Neurosymbolic AI: the 3rd wave the authors talk about coupling numeric and symbolic representations rather than neural and symbolic, which is odd. It would help the clarity of the paper tremendously if the authors started from their intended definition of neurosymbolic AI, present the taxonomy of their choice, and then clearly position their work within this field/taxonomy using the conventional terminology. Translating symbolic knowledge into the network architecture is rather broad and generic.
It would also help if the authors presented NeSy approaches in the section with the same name. Currently, it consists of references and high-level descriptions of objectives of approaches, without providing sufficient details to really provide an overview and a context for the proposed approach. At the end of this section, the authors state that "we are not going to study how to embed a symbolic
representation onto a given numeric space" (3.35), however, earlier they state that their objectives include "embedding symbolic data structures in a metric space" (2.13).
The notion of cognitive science, biological plausibility of solutions, the brain, and consciousness keep on recurring without yet a clear connection to the proposed approach up to page 4. In most cases, the notions are not sufficiently discussed or founded. What is the connection of the proposed approach to cognitive science? How is it biologically plausible?
It is quite unclear why the authors would take a figure from reference 39 without describing the work done in the same reference in the text. Furthermore, Fig. 1 is not anchored, that is, referenced in the text.
The claim "this also clarifies that the numerical aspects pertain primarily to the biophysical side, whereas the symbolic representation relates mainly to the behavioral stage" is not exactly substantiated by the context in which it is raised. I would propose mitigating this claim and also considering that behavior can be highly numeric, e.g. response times, reaction times, etc. I would also propose to provide more context for the first part of the claim that numerical aspects pertain to the biophysical side. At the end of this paragraph, the authors refer to a development proposed in the paper. It is quite unclear what this development might be or refer to.
The idea of modeling brain activity as an ontology is directly presented before talking about how brain processes can be modeled with "such a well-defined framework", suggesting that the provided reference [30] actual performs such modeling with an ontology, which to the best of my understanding it does not. Furthermore, the OWL2RL paper is explicitly given as a good source for a general introduction on ontology. In my opinion, there are a lot of other publications that are truly general introductions to the idea of an ontology and not related directly to a specific language. This request to please consult other references for details on ontology modeling also seems to be randomly placed between sentences in this section.
The notion is raised that a general neural network indeed models the brain, "modeling the brain explicitly using symbolic representations contrasts with the use of very general deep-learning “black-box” tools,", however, the cited reference [57] rather makes the point that a deep learning model for circuits cannot be expected to provide a modeling engine for the brain. The no free lunch principle in that publication is that deep learning models cannot just discover brain-level structures on their own. I would propose clarifying that sentence, mitigating this claim and especially avoiding to present deep learning as universally questionable. The following "What predicts everything, predicts anything." I fail to understand.
The authors state the following:
"We also provide (which is uncommon in scientific papers) a large set of links to help readers from different domains understand unfamiliar terms, primarily via Wikipedia"
I do not quite understand why you would keep such a practice or sentence in a scientific paper, explicitly acknowledging that it is indeed "uncommon in scientific papers". It is uncommon for a reason, as the context for a diverse reader group to understand an article should be given in the article itself and too basic background knowledge may be provided in the form of suitable references. Since it is not a (text) book not all basics should be explained anyhow. It is even worse practice to make the links in the text not explicitly visible, but only noticeable when accidentally clicked on.
Not a lot of what follows in the section entitled "A “natural” usual way to represent cognitive knowledge" relates to cognitive knowledge, but rather to different approaches of representing knowledge in general and with a connection to cognitive science, but not cognitive knowledge. The "natural" usual way is also quite unclear. Maybe a more transparent heading could be considered? In the description of Figure 3, the authors state that concepts are anchored in an input/output setting, that is, stimulus/response. From the figure and current description, it is impossible to understand what the input and the output may be and what the stimulus and response may be.
While the coherence is somewhat better starting from Section 3.2., there is the issue of co-dependence between the data types and the edit distance, leading to having to repeat ideas in both sections. Since these subsections present the very core idea of the submission, it might be worth optimizing the presentation of contents and ensuring argumentative clarity, the latter is already better here than in previous sections.
While the data types string and numeric are widely known, I have not come across the definition of a Boolean type extended by 0 unknown as modal. Is this the author's own not previously known definition? Maybe this could be stated as a proposal rather than a fact or equipped with a reference if taken from previous sources.
The repeated representation of very lengthy and at times formal definitions in the footnote does, in my view not help the readability of the article. Either the formal definitions are relevant and be part of the article, if they are too long they might be placed in the appendix or they are irrelevant and therefore omitted.
The discussion should provide findings, limitations, challenges and relations to prior work, not introduce entirely novel ideas of how the proposed approach is implemented computationally and how efficient this computation is. Instead, the discussion only presents novel aspects of what they wish to contribute.
Minor comments:
3.27 A step ahead, -> Odd and hard to understand choice of linking device
3.44 Since [46] work, -> ?
4.36 At the difference of Marr’s original three implementations, -> In contrast to, Different to...
4.36 three implementations, algorithmic and computational levels, , modern vision, -> three levels of computational, algorithmic, and implementation, a more recent perspective
4.49 In this direction, -> many of the linkers are odd or simply wrong; I do not indicate all of them as considerably too many; please revise thoroughly
5.50 In Figure 3 examples, -> grammar
5.50 an input/output -> an input/output what?
6.32 a bird’s knowledge -> this means the knowledge of a bird, i.e., the knowledge a bird has, which I doubt is easy to model or even investigate and most certainly not what you wanted to say
6.30 This corresponds to the [32] approach, -> this would be so much easier to read if you could just say Conceptual Spaces approach [32]
6.43 quoted before -> why is it worth mentioning that this is not the first time you cite this publication?
6.44 “between” to birds -> two maybe?
6.44 This notion will be correctly formalized in the sequel. -> ? A future publication?
7.20 appendix A.2 -> Appendix A.2
7.27 Given the previous computational objective, -> I am honestly uncertain what this refers back to
7.28 lever -> key mechanism? primary tool?
7.33 Appendix A.2
9.18 makes the job (also before) -> gets the job done? but highly informal and better to revise
9.50 We are in a metric space by virtue of the edit distance. -> rather informal
9.51 editing distance (again later) -> edit distance as in the line before, right?
10.25 b f s_1 / _2 -> What is b f s?
...
13 Figure 7 seems to not be referenced or anchored in the text.
15.23 On a first hand -> The first criterion is that or First,
16.28 quoting -> which are two very ... quoting is not equivalent to citing or referencing
16.48 as closed as -> as close as
22.50 Edition distance -> before edit distance and editing distance
