Reviewer has chosen to be Anonymous

Overall Impression: Good

Content:
Technical Quality of the paper: Good
Originality of the paper: Yes
Adequacy of the bibliography: Yes, but see detailed comments

Presentation:
Adequacy of the abstract: Yes
Introduction: background and motivation: Good
Organization of the paper: Satisfactory
Level of English: Satisfactory
Overall presentation: Good

Detailed Comments:

This is a very interesting position paper that envisions a neural cognitive cycle, combining neural and symbolic computation, to support human-oriented explanations.
The paper overviews some related techniques and methods, and then outlines how these could be used in the provision of a neuro-symbolic cycle from a XAI perspectives, by describing objectives, hypotheses and further challenges.

Overall the paper is well written and easy to read. It provides a nice perspective of neuro-symbolic explainable AI which makes the paper very sensistic for the inaugural issue of the journal.

I have only a few comments about Section 3 and Section 4.

Section 3.

- As far as the objectives are concerned it is not clear (or not well motivated) why Explainability for AI experts and AI stakeholders have been splitted into two objectives. Clearly, the audiences are different and the generation of explanations for these categories will imply to adopt different techniques and methods, nonetheless, O2 and O3 seem to have a shared goal, that is the generation of intelligible explanations for human users.
- Perhaps a graphical representation of the neural-cognitive cycle would enhance the reading.
- Some parts of the text are a bit repetitive.

Section 4.
- It seems to me that H1, H2, and H3 more than hypotheses can be considered enabling techniques, methods and expertises that can be leveraged to support the neural-cognitive cycle discussed in Section 3.

Section 4.1
- What is ILASP?

Minor points:

- pp.2, line 40: [12]or -> [12] or
- pp.2, line 40: [13] ; -> [13];
- pp.3, line 3: [19][20][21] -> [19,20,21]
- pp.4, line 11: compositionality. as -> compositionality as
- pp.4, line 24: known[32] -> known [32]
- pp.4, line 31: and and -> and
- pp.5, line 4: [36][33][37] -> [36,33,37]
- pp.5, line 17: [[25] -> [25]
- pp.5, line 46: the concept of AUC has not been introduced.
- pp.6, line 6-7: [41][42] / [25][26] -> [41,42] / [25,26]
- pp.6, line 14: feedback The -> feedback. The
- pp.6, line 15: should encodes -> should encode
- pp.6, line 30: dom heuristic -> domain (?) heuristic
- pp.6, line 33: layer . Quality -> layer. Quality
- pp.6, line 26: In what follows The remainder of this section -> The remainder of this section
- pp.7, line 46: COMPAS 2 -> COMPAS2 (footnote); also, add last access date
- pp.7, line 47: bad 2 -> bad3 (footnote); also, why not citing it as a reference?
- pp.8, line 1: image 4. -> image.4 (footnote); also, why not citing it as a reference?
- pp.8, line 3: report 5 -> report5 (footnote); also, add last access date
- pp.8, line 19: Advances discusses -> Advanced discussed
- pp.8, line 33: sare known -> are known

References:
Please check references, in particular:
- [3]'s author list should be fixed
- Reference [27] should be corrected, the article was published in the AI Magazine in 2022, and not in 2020.

Reviewer has chosen to be Anonymous

Overall Impression: Good

Content:
Technical Quality of the paper: Good
Originality of the paper: Yes
Adequacy of the bibliography: Yes

Presentation:
Adequacy of the abstract: Yes
Introduction: background and motivation: Good
Organization of the paper: Satisfactory
Level of English: Satisfactory
Overall presentation: Good

Detailed Comments:

The paper presents a persuasive argument for the Neuro-Symbolic Cycle in the context of Human-Centered Explainability. It serves as an intriguing position paper that proposes a neural cognitive cycle, effectively merging neural and symbolic computational methods to facilitate explanations centered around human understanding.

The paper provides an overview of relevant techniques and methodologies, subsequently delineating their application in establishing a neuro-symbolic cycle through the lens of Explainable Artificial Intelligence (XAI). It details the objectives, hypotheses, and the various challenges that may arise.

The paper is commendably clear and fluid in its writing style. Some minor suggestions for improvement are outlined below.

Conceptual Framework and Methodology:
- The paper introduces a compelling conceptual framework for integrating neuro-symbolic approaches in explainable AI. However, it would benefit from a more detailed discussion on the methodologies for integrating these two paradigms. Specifically, how will the symbolic reasoning components interact with the neural network layers, and what are the anticipated challenges in this integration?
- A deeper exploration of the data types and structures that would be most suitable for this neuro-symbolic cycle would be beneficial. For instance, how will the proposed cycle handle different types of data (e.g., structured vs unstructured data)?

Evaluation and Validation:
- The paper would benefit from an outline for the evaluation and validation of the proposed neuro-symbolic cycle. What metrics or criteria will be used to assess the effectiveness and accuracy of the explanations generated?
- Including a discussion on potential benchmark datasets or experimental setups that could be used to validate the proposed approach would add considerable value to the paper.

XAI:
- The paper could benefit from a deeper exploration of how neuro-symbolic integration can address specific challenges in explainability, such as dealing with complex, high-dimensional data.

- The paper would benefit from a more consistent formatting style for citations and footnotes.

Reviewer has chosen not to be Anonymous

Overall Impression: Good

Content:
Technical Quality of the paper: Good
Originality of the paper: Yes
Adequacy of the bibliography: Yes

Presentation:
Adequacy of the abstract: Yes
Introduction: background and motivation: Limited
Organization of the paper: Satisfactory
Level of English: Satisfactory
Overall presentation: Good

Detailed Comments:

First of all, very interesting paper addressing explainablility using neuro-symbolic methods of DNNs with bringing humans in the loop, particularly with training a DNN with injected prior knowledge. I am very interested in reading about the experiments and their results. Please see my detailed comments below.

Introduction:
There seems to be a missing link between argument against 'model-based interpretability' and argument for the proposed neuro-symbolic cycle. For example, how is the proposed neuro-symbolic cycle can overcome the limitations of incomplete and incorrect explanations? Will the proposed cycle able to produce both high accuracy preidctions and trustworthy explanations? In the given example (attribution maps), is it the case for all post-hoc and interpretable methods? As in do all XAI methods highlight a portions of an image that are responsbile for a predicted outcome without specifiying how the outcome was produced?

State of the Art:
Perhaps some citations for for sentences from line 35-38? E.g., post-hoc approaches and attention-based approaches for medical image analysis.

Key Elements of the Extraction-Explanation-Injection cycle:
- H2: the assumption here would be that the knowledge extracted from the trained GNN or other networks will be treated as ‘the absolute truth’. Although this extracted knowledge will be validated by the medical professionals, but how will this ensure the explainable model to be 'robust' and 'accurate' given that different optinions might be given for the validation process, especially in diagnosis?
- Continue on from the previous point, what will happen if complete opposite opinions/knowledge are given by the experts during the validation process, will both be injected and propogated into the non-trained network and will both be given equal weights?

Some minor presentation comments:
- What are 'KGs' in line 2 under section 4.3?
- The uses of brackets might be better to replace with commas or full stops.