Reviewer has chosen not to be Anonymous

Overall Impression: Good

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

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

Detailed Comments:

The paper presents a novel neuro-symbolic approach called Ness-\pi that is able to:
1. Take as input complex scenes
2. Invent new predicates that aid in the classification of complex images

Strengths:
- The paper is very relevant to the community
- The experimental analysis is very detailed
- The solution proposed is interesting

Questions:

While reading the paper, the below questions arose that could contribute to enhancing clarity in the revised version of the document.

- Is it fair to talk about predicate invention? Personally, I find this name a bit misleading. Indeed, initially, I thought the system was inventing all the predicates (including the base ones like shape, colour, in, etc. etc.). When I got to the description of the experimental analysis I then understood that the system can only invent new predicates that represent more complex concepts than the given ones. For this reason, wouldn’t it be better to talk about “learning complex concepts”?
- Can the system be trained end-to-end? Or each module is trained as a separate component?
- How much complexity does the 3D add? It seems that there is not much difference between handling 2D images and 3D inputs.
- The proposed system seems to be very dependent on the initial set of predicates. Have you tried any experiment where the initial set of predicates is different? It would be interesting to study what happens when either some predicates are removed or some redundant ones are added to the ones listed in Table 1.

Suggestions:
- I would rewrite the initial part of the paper to include a running example drawn from your experiments to explain exactly the capabilities of the system. Indeed, initially, I was confused about how the authors were training the Mask-RCNN without labels, what was the depth map and how the grouping module worked. With a running example, it would be possible to clarify: (i) what are exactly the outputs of the Mask-RCNN, (ii) what are the inputs and outputs of the grouping module etc. etc.

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 overall paper is easy to read and follow with clear explanation of the entire NeSy-π approach. There are clear evidence of their contribution by utilizing a NeSy approach to invent new predicates that can improve the reasoning of prediction from 3D Visual Scenes.Good amount of description on the method and sufficient experimental examples to illustrate their evaluation.

The following are some comments that may be useful for further clarity.
(1) It may be better to highlight that NeSy- π refers to the NeSy-PI component in Fig. 2. It is abit confusing when the term 'NeSy- π' system is being used in Section 3 and NeSy-PI as a concept inventor in Figure (also in pg. 5 line 3 where NeS- π receives ....).

(2) Is P-Measure the percentage of positive examples or the aggregated average of the confidence values (as given by eq. 1 and relating to the definition of rc in Pg.5 line 43-44)? Similar to the N-Measure.

(3) Is there a reason for using P.N measure over other conventional evaluation measures?

(4) Given the complexity of using different components within the workflow, will the quality of Mask_RCNN affect the object perception and other downstream steps? Is there any vetting process to ensure any error is not cascaded down the workflow?

(5) Similar to (1), it is abit confusing to understand Algorithm 1 includes without a clear definition of what NeSy- π refer to.

(6) Fig 7 and other example figures in Appendix B: it will be useful to state what the value of rho is in the left images. While it is clear that there is a clear difference in distance, it may be clearer to understand whether the model has learned the new invented predicates.

(7) Pg 19 line 32: Typo 'firs' should be 'first'