Introduction to the E-Sense Artificial Intelligence System

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The paper suggests a three-level cognitive architecture (ontology, memory, neural layers) attempting to integrate ideas from statistical modeling, ontologies, and cortical function. The attempt to align computational structures with biology is an interesting concept. Aligning with columnar cortical organization and borrowing from Gestalt psychology as well as ordinal learning gives new, if hypothetical, insight into cognitive modeling. The ordinal learning idea of a capability to reproduce order sequence is an exotic algorithmic facet not so emphasized in the normal AI mainstream science today. Significant limitations are indeed encountered by very many of great magnitude, necessitating much higher improvements;

• Normal scientific rigor of the work here is not matched by this paper. Models lack definition, there is no sharpness to measurement, and also the approach taken to evaluating anything is really, really subjective.
• The paper mixes together implementation facts, authorial speculation, and philosophical musing without explicit demarcation. This makes a huge sacrifice in terms of clarity and academic credibility.
• Weak experimental evidence underlies the paper. Experiments are tried on extremely small toy instances (e.g., sorting cooking recipe instructions), with no comparison against baseline or state-of-the-art methods in an effort to estimate performance or usability.
• Key terms such as "functional structures," "type-based clusters," "neuronal types," and "ordinal learning" are used without uniform or proper usage. This discourages replication and misleads the reader in technical description.
• Most of the referenced citations (especially those underpinning main arguments) are to the author's past work. Outside verification from peer-reviewed sources is minimal, reducing the objectivity and ground of research.
• The article is repeatedly describing the model's beneficial usage as nebulous (e.g., "an application for its functionality is not yet clear"). This prompts us to ask how publishable it is in a scientific journal.
• Figures are cited but poorly integrated into the explanation. This makes the paper difficult to follow and lowers its professional presentation.

Comments on the Quality of English Language

There are numerous grammatical issues, sentence fragments, and loosely organized paragraphs in the manuscript.

Author Response

1. Thank you, some statements are now more definite. The architecture has been rigorously designed with the consideration of many other works, often for specific points. The ordinal testing simply shows that it can impose that order on new content. The ontology testing is more subjective. I have made efforts to convey the ideas in a clear manner. I have highlighted the main changes, where a new section added to the Introduction lists potential advantages over the current models. Also in the related work section. 

2. I am able to understand the design because I am familiar with it. If there are still terms that are not well understood, if you can list them I will address that problem.  I think that there is mathematical consistency throughout the whole design.

3. The experimental results are subjective, but the paper is mostly about an architecture. A new section in 4.4 makes a comparison with a word vector, but the results for that are equally vague and that is the established technology. However, a new conclusion has been made. I think that more rigorous testing should be done in a separate paper.

4. 'Functional structures' and 'type-based clusters' have only 1 meaning in the paper. The functional structure is the structure that makes up the function. Type-based clusters is clusters based on type. 'Neuronal types' does not now exist. 'Ordinal learning' is defined in the abstract, with that meaning remaining throughout the paper. I think the reviewer has hallucinated here.

5. It is the case that the architecture is very new and so there is not much reference from other researchers, apart from an extensive related work section.

6. It is only stated once. There is a lot of scientific justification to the paper.

7. Thank you. An additional description for figure 1 has been added.

Reviewer 2 Report

Comments and Suggestions for Authors

The authors in this manuscript propose the E-Sense Artificial Intelligence system, which comprises of a memory model with 2 levels of information and then a more neural layer above that.

This manuscript needs major revision, and the authors need to address the following issues:

1)While the paper introduces a novel three-tiered architecture (E-Sense), the comparative advantages over existing AI systems should be more explicitly articulated. Specifically: a) What key limitations of current approaches (e.g., reliance on weighted statistical models in LLMs, rigidity in sequence learning, or high data dependency) does E-Sense address? b) How does the unweighted Markov-based memory model or ordinal learning mechanism provide unique benefits compared to conventional deep learning or symbolic systems?  

2)In the Introduction section, the authors’ research motivation is not clearly stated. Suggest the authors Explicitly summarize the unresolved gaps in contemporary AI, such as: a）The inability of purely statistical models (e.g., GPT-4) to handle dynamic reordering of learned sequences. b) High computational costs of training large-scale weighted networks. c) Limited biological plausibility in modular brain-inspired designs.  And then connect these gaps directly to E-Sense’s design goal.

3)The empirical evaluation is currently limited to qualitative demonstrations (e.g., Appendix B) and subjective clustering evaluations, without systematic quantitative validation.

Author Response

1. Thank you for this point. I have highlighted the main changes, where a new section added to the Introduction lists potential advantages over other models. There is also new content in the related work section.

2. This has been addressed in the Introduction and in the Related Work sections.  I think that there is mathematical consistency throughout the whole design.

3. The ontology testing in section 4.4 has been expanded with further descriptions and a loose comparison with a Word Vector result. A new conclusion has been made. The ordinal testing (Apendix B) is just intended to demonstrate that it works.

Reviewer 3 Report

Comments and Suggestions for Authors

The manuscript proposes a novel three-layer cognitive model called E-Sense, which attempts to bridge memory structures with functional neural analogies. While the ambition behind the model is commendable, the overall contribution remains conceptually vague. The paper tends to lean heavily on speculative analogies with biological and cognitive theories, but often lacks rigorous justification or empirical validation to substantiate the claims.

The organization of the content is at times difficult to follow. Terminology is introduced without formal definitions, and the reader is frequently expected to infer the intended meaning. For instance, the transitions between memory levels and their purported cognitive relevance are presented descriptively but not analytically, which hinders clarity. The narrative is also burdened by repetitive statements and assertions that lack experimental backing, particularly regarding the role and structure of the "ordinator" and how it truly achieves ordinal learning. This key mechanism is described at length, yet remains algorithmically ambiguous and unconvincing in its necessity over simpler sequence modeling techniques.

While the manuscript attempts to provide a number of comparative references to biological phenomena and existing AI systems, these comparisons are often superficial. References to cortical columnar structures, heavy-tailed networks, and Gestalt psychology are loosely mapped onto elements of the model without formal modeling or empirical studies that would support such mapping. This creates the impression of post hoc rationalization rather than hypothesis-driven design.

Moreover, the test results presented in the appendices are rudimentary. The evaluation methodology is informal and lacks quantitative benchmarks or comparisons against existing techniques. The use of a few manually inspected examples is insufficient to demonstrate the scalability, reliability, or generalization capacity of the proposed system. The reliance on subjective interpretation for validating clustering or ordering further weakens the strength of the conclusions.

To enhance the scientific quality and technical depth of this work, the author is strongly encouraged to study related works that provide concrete, well-evaluated approaches in neural modeling and AI systems. In particular, “Towards model-free tool dynamic identification and calibration using multi-layer neural network” presents a grounded and mathematically sound neural architecture that could serve as a methodological reference. Additionally, “A Review of AIoT-based Human Activity Recognition: From Application to Technique” offers insights into integrating cognitive modeling with application-driven AI systems, which could help refine the functional relevance of the proposed framework.

Author Response

1. I have highlighted the main changes, where a new related work section adds more support to the idea that the design closely relates with a biological model. The architecture has been rigorously designed with the consideration of many other works, often for specific points.

2. I have made efforts to convey the ideas in a clear manner. Because most of it is not standard, the ideas will be new to most readers. The ontology is intended to produce the biological mechanism of recognising types, but it is not intended to be a biological structure. The upper neural layer looks to be structurally more biological. The document is already quite long and it is not possible to add even more detail. That would require separate papers. For example, with regards to the ordering algorithm, it might be possible to simply assign a number to each sentence in a text and then do a best word count for that sentence, to place some other sentence in that position. But I think that it would spoil the paper to add this type of test, because it requires further descriptions that might become too focused and it is not quite the same. It is the idea behind the new algorithm that is important.

3. Superficial, but they still clearly exist. OK, the Gestalt numbers could be anything, but still match with the architecture. The top level columns have some biological similarity to cortical ones because specific neurons are also generated. While the design has followed some consistent themes, the biological associations have been found afterwards. As this is from a computer science perspective, I consider that to be supportive and not detrimental.

4. I have added more information to the ontology testing, where a new conclusion has been made. The ordinal testing is simply intended to show that it works. I think that there is mathematical consistency throughout the whole design.

5. More comprehensive results would be very taxing and I think, require new papers. It is mostly about a broad architecture and is not as focused as the paper that you suggest. I think that the topic of those papers is slightly different. I was only able to download 1 of them. If I can, I will consider the second paper when I develop the functional model further.

Reviewer 4 Report

Comments and Suggestions for Authors

This is a well-written article describing the E-Sense artificial intelligence system. To further enhance the value and significance of the article, the reviewer believes it would be desirable to expand the section that discusses comparisons with other columnar structures in the neural cortex, particularly the visual cortex, where similar structures have been identified that respond selectively to stimulus variations in binocular disparity.

Author Response

Thank you for your supportive review. I have highlighted the main changes in the text. I can recognise some neuron types, such as pyramidal, or ones with larger dendrites, but I am not very expert in the biological scoences. I was hoping to obtain more substantial results before trying to make these types of comparison clear.

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The revised submission offers enhanced organization, especially in the explanation of the three-tiered design. Information is logically ordered better and figures are employed to enhance comprehension.

The concept of "ordinal learning" is better explained and with stronger examples as supporting evidence, so the distinctiveness of this learning mechanism amongst other AI learning mechanisms is easier to grasp. The process by which sequence order is identified and realigned is better explained and supported through small test examples.

The biological analogies, particularly to forms of neurons (unipolar, bipolar, pyramidal) and cortical column structure, are now more systematically presented, further establishing the theoretical foundation and categorizing the model as biologically inspired but not statistically strict.

The related work section has been enriched with references to bleeding-edge developments such as transformers, neuro-symbolic systems, and Hopfield networks, providing valuable comparative context. But little else remains to be done to further polish the paper;

• The testing remains largely qualitative and subjective. While acceptable for proto-models at early stages, adding even basic quantitative measures (e.g., precision, recall for cluster precision) would significantly enhance scientific rigor.
  Informal introductions to key terms such as "typing," "ordinator," and "unit of work" remain. A glossary or more frequent formal definitions would make reading and replication easier.
• The article admits that "a use for its functionality is still not clear" and the system is less precise than cutting-edge systems. An addition of a section of potential uses (e.g., compression of memory, symbolic interpretation in low-data regimes) would address this and align with journal requirements.
• There remain a few grammatical errors, typographical irregularities (e.g., "its'" instead of "its"), and awkward sentences to make it difficult to read.

Author Response

4. The related work section has been enriched with references to bleeding-edge developments such as transformers, neuro-symbolic systems, and Hopfield networks, providing valuable comparative context. But little else remains to be done to further polish the paper.

Thank you.
4a. The testing has been moved to a new section 7. A very basic precision test has been added to the ontology testing.
4b. A description of the 'type' is given on page 2. Ordinator is just the name used for the ordinal learning program. 'Unit of Work' is defined at the start of section 4.4.
4c. The sentence 'functionality is still not clear' has been reworded. The uses that you mention are indeed true.
4d. Thank you, I have made more corrections.

Reviewer 3 Report

Comments and Suggestions for Authors

1. While the paper presents an intriguing biologically-inspired model, the exact problem it solves remains vague. The manuscript should clarify whether E-Sense is intended as a cognitive simulation, a practical AI tool, or both.

 

2. The separation between the three levels—Markov-based memory, ontology, and neural functionality—is foundational, yet their interactions lack formal definition. How data flows and transforms across levels should be made explicit.

 

3. The ontology construction is presented as a novel contribution, yet the methodology remains underdeveloped. The “ensemble to tree” transformation is described vaguely and would benefit from algorithmic or pseudocode specification.

 

4. The concept of “ordinal learning” is novel but insufficiently contrasted with existing approaches such as sequence modeling in LSTMs or transformers. The claim that this approach is not implemented elsewhere requires stronger literature support.

 

5. Despite frequent comparisons with cortical structures, the neurobiological plausibility of E-Sense is asserted more than demonstrated. The references to neuron types (unipolar, bipolar, pyramidal) feel analogical rather than rigorous.

 

6. The testing methodology is underpowered. Demonstrations with children's stories and cooking instructions do not offer a convincing benchmark. Empirical validation on established NLP or cognitive datasets is necessary.

 

7. While orthogonality is emphasized as a design principle, its implications for generalization, scalability, and robustness are unclear. This needs to be quantified or at least examined under adversarial or noisy input.

 

8. The figures (particularly Figure 2 and Figure 3) are not referenced or explained thoroughly in the main text. They should be better integrated and annotated to help the reader follow the architectural logic.

 

9. The distinction between this work and previously cited architectures like SPAUN, SOAR, and GPT models needs clearer articulation. What measurable advantage does E-Sense offer over these models?

 

10. The language occasionally drifts into speculative territory (e.g., the role of “heavy-tailed neurons” in ordering), which weakens the technical rigor. These ideas should either be substantiated or labeled clearly as hypotheses for future exploration.

 

11. The role of the associative network is described as “not yet implemented,” which weakens the system’s claims of completeness. The paper would benefit from describing implementation plans or experiments underway.

 

12. Though the paper is rich in ideas, the lack of a running example or application scenario makes it difficult to grasp the model’s practical use. A concrete use case, even speculative, would improve accessibility and relevance.

 

Author Response

1. New text in the Introduction should answer this. The usefullness of the new algorithms outside of this model is a bonus, where the overall cognitive architecture is the primary purpose of the work.

2. I think that this is answered in section 4.1. The Lower to middle transition is well defined, but middle to upper still requires additional work. Some new information has been provided in sections 4.3 and 7.1.

3. Again, the algorithms are too complicated. But section 4.4 gives some more descriptions.

4. The paper should not be evaluated as the same as sequential methods. A new senence has been added to the start of section 6. An extensive literature review has been carried out.

5. It is only from some preliminary results and so I cannot be too definite. I have added a new section 8.2 to describe the ideas about the cortical columns, but it would only be an analogy anyway.

6. The testing has been moved to a new section 7. A very bsic precision test has been added to the ontology testing. Again, only preliminary results.

7. I think that this would require a new paper with more extensive testing, but it would be a good basis for further testing.

8. I think that these 2 figures are quite well used in the text, but descriptions have been added to each Figure listing, thank you.

9. The relation to the other models in related work has been made clear in the conclusions. Comparisons with word vectors from ChatGPT is also clear and also differences with neural networks.

10. Thank you, that has been corrected in section 6.

11. Thank you, that has been corrected in section 4.4.

12. I hope that the reader can follow the flow of information and understand that it is a cognitive model. I think that an application scenario will follow later, after the model is fully constructed.

Round 3

Reviewer 3 Report

Comments and Suggestions for Authors

The current version can be accepted.