<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Ekneumann on Medium]]></title>
        <description><![CDATA[Stories by Ekneumann on Medium]]></description>
        <link>https://medium.com/@ekneumann?source=rss-6b577461748d------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*iLlvxYk8ICkmVYvu</url>
            <title>Stories by Ekneumann on Medium</title>
            <link>https://medium.com/@ekneumann?source=rss-6b577461748d------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 09:49:18 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@ekneumann/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Ontologies aren’t what they used to be… actually, the world has changed]]></title>
            <link>https://medium.com/@ekneumann/ontologies-arent-what-they-used-to-be-actually-the-world-has-changed-397162641eea?source=rss-6b577461748d------2</link>
            <guid isPermaLink="false">https://medium.com/p/397162641eea</guid>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[yoneda-embedding]]></category>
            <category><![CDATA[ontology]]></category>
            <category><![CDATA[knowledge-graph]]></category>
            <category><![CDATA[simplex-semantics]]></category>
            <dc:creator><![CDATA[Ekneumann]]></dc:creator>
            <pubDate>Mon, 01 Jun 2026 03:00:49 GMT</pubDate>
            <atom:updated>2026-07-21T14:44:01.778Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*NXsXbXacJ6yTI7tKgu2KNg.jpeg" /></figure><p>Ontologies are powerful and in the 2000’s they were a cutting-edge idea, but they were challenging to explain — most informaticists expected them to be identical to database schemas just with more semantics… they are not. They are an abstraction different from what most software developers and informaticists had been trained on. They were supposed to provide a bridge between the software and biomedical researcher communities. As it turned out, concept semantics took second seat to classifying taxons, like species and gene hierarchies. This was exemplified by the creation and use of the Gene Ontology (GO), which was launch in 1998. GO is actually not a formal ontology, yet it was more readily usable by bioinformaticians as a means to organize emerging genomes from large-scale sequencing efforts, more than any other ontological system. I was present at the ISMB’98 meeting in Montreal where ontologies in biology were first proposed as well as the 98-TIGR meeting eating pizzas with Greg Helt and Michael Ashburner ( former drosophilist colleagues) while they began codifying a GO draft in a spreadsheet. The late Michael Ashburner was present at both, and took issue with the formal definition of ontologies as all descending from a top concept like “Thing or Entity… or God” as he put it. Back then, it was how best to move from closed-system of academic tools to an open model for knowledge exchange [1].</p><h3>When are ontologies useful?</h3><p>To be fair, abstractions are powerful and re-usable, but need to align with the kinds of activities and associations with which the community is currently pre-occupied . The biomedical community needed a classification code much more than how to relate codeable biological entities as class objects. The two are related, but the former is closer to a tagging system, and easier for biologists to understand in the world of spreadsheets: a column, not an object defined in java.</p><p>There is another wrinkle to briefly mention involving choice of ontological organization based on how we intend to use them: What is the basis for our classifications? Is it by biological functions, by evolutionary continuity, or by logic principles such as information storage/replication and computational abilities (e.g., adaptive immune response)? Gene Ontology addresses the former, though things have been labelled with how we currently believe biology works (functionality is not only what disease a gene prevents, but all it is supposed to do during normal development and homeostasis — surprising to some, we do not yet have a complete understanding of this). And since biology relies on the gene storing the blueprint to make a operational protein, how do we ontologically label the storage part, as related to the regulation and function of the expressed part? I will hold off digressing for now, but it is an open area not yet addressed by today’s ontologists.</p><p>If you were studying the oncogene ERBB2 (a cancer driver when mutated) as well as assaying its activity as a protein, both are easily labelled with the gene ontology term, GO:0038128. If you had defined genes and proteins distinct java (coding) objects of type “gene for ERBB2 signaling pathway” and “protein for ERBB2 signaling pathway”, you’d still end up relying on the GO taxonomy, but it would have been extra work with little gain. As far as specific concepts go, there is little practical gain making object classes that are built from “signaling pathway” or “gene regulatory” classes. So the rationale to organize code using ontologies makes little sense aside from using them as a type-labelling system for soft-typing. That was then, and the ontology-code angle is not pursued much anymore…</p><p>On an emerging front, as we now try to reboot knowledge representations with knowledge graphs, and attempt to associate formal semantics with LLMs, ontologies may have a new-life. The structures can map into sufficiently augmented LLM models, and this can empower models to have more applications in regulated environments. However, we have also learned a few things along the way that we should acknowledge so as to streamline efforts. If hierarchical classifications, aka <em>tree graphs</em>, are central, we need to ensure they do not get in the way of <em>general graphs</em> that encode diverse relationships. There is even a real question whether concept hierarchies or subsumption logic in graphs need to be hardwired (into concepts) if taxon labels already exist and if there are many more relationships between entities or concepts than sub- and super-classes.</p><p>Classifications models (aside from GO taxonomy) have mostly been created by non-biologists, and having observed this for a few decades, I don’t believe this has gotten us very far. This needs to be corrected, and at its root it needs an alternative organizing principle. There is currently a strong drive to repackage computational tools that until recently were called <em>machine learning</em>. Terms like intelligence and knowledge have been inserted and re-defined to promote both awe and fear. Maybe it will continue to grow monotonically into a new technical and societal revolution, but it could also run out of steam if the marketing hype used to drive venture value have promoted it beyond it’s current abilities. I hope to address the issues a bit around a better grounding for knowledge representation.</p><p>Formal ontologies as a design tool are notoriously difficult to wield. They are often: subjective, rigid, difficult to maintain or adjust to changing business needs. They are often referred to as a Time Trap, and difficult for developers to take full advantage of them. For the amount of effort of engineering some semantics, they are often biased and limited in utility. They are often descriptive from one point of view, and not insightful or prescriptive in general. LLMs to a large degree already provide more utility, so perhaps ontologies may need to recognize these benefits to become more practical themselves. LLMs are sometimes used to <em>extract</em> conceptual forms around key concepts in a domain to help guide ontological definitions. This still requires a human to prime the pump so to speak.</p><p>Specifying typed nodes arises from our classification view of ontologies rather than something that is based in reality. Does it arise from nature, from human perception, or from how the science community currently views the world? Most believe it’s all three, but it’s certainly an approximation, a model that is always wrong to some degree. It’s one kind of “embedding” one can apply, but probably should not be hard-coded. As I stated earlier, knowledge representations are on the rise, and things seemed to be gaining general interest due to the many applications of LLMs.</p><p>LLMs are also showing us that numeric embedding work as well if not better than forcing a universal tree to fit, by including multiple kinds of non-taxonomic relations when aligning tokens with implicit concepts. The inclusion of all forms of relations via embeddings is what makes LLMs so effective, surpassing the capabilities of original formal ontologies. The lesson is: all relations between words/concepts is more powerful than any taxonomic-based structure based on types. The breadth of relations form a natural bridge between deeper multi-word concepts, and this hints at a different kind of mapping for knowledge graphs as well. For me, the intersection of semantic shapes and LLMs is an area rich with many applications.</p><p>As an alternative to high-dimension embeddings, weighted graphs have been shown to have an equivalence with token embeddings. These weights can be over local sections of graph. From a Knowledge Graph (KG) point of view, this aligns conveniently with the use of <em>shapes</em> <em>of relations</em> between nodes in a KG, i.e., semantically consistent patterns or subgraphs. There’s a lot of power around subgraphs.</p><h3>Yoneda Embeddings</h3><p>Subgraphs are more than small sets of nodes and edges, in particular, they encompass a higher graph logic frame known as <a href="https://en.wikipedia.org/wiki/Yoneda_lemma"><em>Yoneda embeddings</em></a>.<strong> </strong>Instead of looking at an object in isolation (<em>𝑪</em>), the Yoneda embedding studies how all other objects in the category “see” it (<em>𝑪→𝑫</em>). Yoneda has been rigorously proven to be a <em>full and faithful mapping of objects</em>, meaning it preserves all the properties and connections between objects or concepts exactly as they were in the original ontology or schema. Yoneda is famously summarized by the motto: “Tell me how you relate to everything else, and I will tell you who you are”. Formally, you can collect all the relations pointing from object <em>𝑪 </em>to <em>𝑫,</em> and now have <em>𝑪</em> point to these relations (<em>𝑪→𝑫</em>), what we might refer to as <em>subgraphs</em> or <em>property graphs</em>. This can be extended to all relation before and after <em>𝑪 </em>, so that rather than referring to <em>things</em> as nodes, it now applied to edges, hyperedges, and mappings in general. The mathematical term is <em>Functor</em>, but we won’t worry about that now<em>. </em>One can think of all this as a precise, high-resolution translation of formulas and knowledge into a universal language.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/828/1*qRdCV1RdDE3qzh7-el3Asg.jpeg" /><figcaption>Fig 1. Yoneda embedding via Functor hₓ, to map objects in U to all relations (maps) between U and X.</figcaption></figure><p>Both ontologies and LLMs attempt to aspire to Yoneda principles, by trying to connect everything in form of completeness, but the implementations are not equivalent. For example, even though relations are key, the complex set of relations around an entity <em>𝑪 </em>does not usually receive its own identifier that can be listed. The introduction of <em>shapes</em> changes this and makes relations with neighbors key in defining a Concept. Rather than having to know an objects type, the type <em>is</em> defined by the shape constraint of its essential neighbors. It directly affects concept look-up, and shifts representations like OWL, into one requiring less <em>ontologist-blessed concepts</em> and more on how the entity is to be situated with other essential linked concepts¹. This is best defined by domain users of these concepts, for instance, genes by geneticists or biomechanisms by molecular biologists. This of course raises questions of how complete a <em>shape</em> needs to be, which I will return to shortly. Some research has already been performed that formally links Yoneda to formal ontologies (Seremeti L, Kougias, 2013 [4]).</p><p>The <em>shape </em>approach is closely tied to topological data analysis (TDA) [2], which also relies on the neighborhood of data points. Since TDA is used often in data science, one is able to merge concept-shape integration with data analytics directly, and bring in LLM results into the same composition. In each case, labeling the shapes is just as important as having labels for the entities or data points. As you’re probably guessing, this provides a reasonable approach to <em>Neuro-Symbolic AI</em> as well, and why labeling a <em>shape</em> is fundamental.</p><p>Implementations of shape semantics have been emerging, spearheaded by the W3C Shape Constraint Language, <a href="https://www.w3.org/TR/shacl/"><em>SHACL</em></a>. It was created for use in complex queries, but it has opened the door for much more. Many ontologists are using <em>SHACL</em> with or in lieu of <em>OWL </em>[3], it really depends on where the community wants to take <em>shapes</em>. One can assume shapes map to <a href="https://www.w3.org/2004/03/trix/">NamedGraphs</a>, another W3C recommendation where, partial graphs over a context receive a label or uri. This can be used to form the basis of shapes on top of shapes. This of course requries some deeper thinking about the best form of semantics so that things don’t go the wild west route.</p><h3>Simplex Semantics</h3><p>The most direct and universal semantic for shapes or shapes of shapes n graphs is leveraging the notion of a <a href="https://en.wikipedia.org/wiki/Simplex"><em>Simplex</em></a>. A simplex is a tiered system of subsets, and a simplex of order <em>k</em> (<em>k-simplex</em>), by definition contains all subsets of it within it as k-1, k-2, … 0 simplices (the “0” term is only the vertices). For example, 2 entities or vertices (<em>0-simplex</em>) can be included in an edge (1<em>-simplex=vertices + edge</em>); 3 vertices can be included as a face with 3 edges (2<em>-simplex=3 vertices + 3 edges + face, </em>aka<em> hyperedge</em>); this continues up to a simplex that holds <em>n</em> vertices (<em>n-1-simplex</em>), written as <em>[n]</em>. The number of instances for each sub-order follows from Pascal’s triangle, that is, combinatorics of subsets.</p><p>Why this excursion into mathematics? I feel this offers a useful, possibly optimal, organizing principle for data, schemas, and ontologies. If one shifts to transforming <em>edges</em> and <em>faces</em> into <em>NamedGraphs</em>, a natural semantic pops-out: all vertices and inner-edges acquire a conjunctive semantic, meaning all inner entities and relations must be <em>TRUE</em> for a set of data under the same context. As a quick example, imagine a data model for a health system that includes doctors, patients, and diseases (just 3 for now). A <em>doctor-patient</em> relationship is a typical starting point, but what happens if it is to be organized by disease (at a given point in time)? In addition to the <em>doctor-patient</em> relationship, one may include a <em>doctor-disease</em> relationship edge (specialist), or a <em>patient-disease</em> relationship edge (diagnosis). All three relationships can <em>triangulate</em> the patient and doctor with the specific disease the patient has been diagnosed by the doctor, and is awaiting treatment (possibly a fourth inclusion). The triangle face is a higher-order 3-relation for a diagnosis (or a treatment): <em>doctor-patient-disease</em> all need to be included in a semantic entity, since it is the conjunction of all 3 2-relationships. One can easily include another typical health data element here, namely, the <em>visit</em> — I’ll leave you to do this on your own. That triangulation itself becomes a labeled entity, which internally includes 3 labelled entities and 3 labelled edges.</p><p>The traditional notion of source and target, or subject and object, break down here, as it now needs to handle higher-order relations. The more you look around, the more examples you’ll begin to identify. In fact, it turns out relational databases (RDB) handle this already while today’s knowledge graphs struggle with it: just as you can define a cross-table between 2 tables, and include properties (which is what a property graph literally reifies), so can you define a table with 3 or more foreign keys to define hyperedges. It’s just that KGs have not yet gotten around to figure out the semantic constraints and indexing ( Neo4J ;) ). What’s more, a RDB can also store properties within the hyperedge, as shown here:</p><pre>CREATE TABLE diagnosis (<br>    doctor_id integer REFERENCES doctors(doctor_id),<br>    patient_id integer REFERENCES patients(patient_id),<br>    disease_id integer REFERENCES diseases(disease_id),<br>    diagnosed_on timedate,<br>    coverage text, <br>    PRIMARY KEY (doctor_id, patient_id, disease_id)<br>);</pre><p><em>Simplex semantic</em> works well with many dataframe systems that have many kinds of relationships. The real question is, how far does one allow these simplexes to proliferate? The simplest answer is, only when necessary and practical. With the above <em>doctor-patient-disease </em>example, this already is encoded in most health datasystem tables as multiple foreign keys, and can be used to generate join tables on demand. RDB systems can use a multi-key search to find all entities, or pairs, or three-way associations.</p><p>For designers of knowledge graphs, it requires a shift in mindset from <em>edge-path</em> solutions (only node to node) to more complete groups. Important to note, that edge-path collections of multiple entity nodes require a biased perspective from one <em>enforced linear semantic</em> view. For a set of neighboring <em>n</em> entities, the total number of full paths through them is <em>n!</em>, which translates into <em>n! </em>ways to define path semantics. This is not a very practical approach to rely on for ontological consensus. This translates into having to choose using one of many possible semantics for multiple entities, and is often why different ontologies diverge and are incompatible. The <em>simplex shape</em> approach collapses this to only one universal group semantic, where multiple query paths would work, since all possible paths between <em>k-</em>related entities are included in a simplex (convince yourself of this).</p><p>How should one determine the big enough shape that captures most local semantics, yet does not become too unwieldy? In other words, what is the Maximum Common Semantic Simplex (MCSS) formed for a given cluster of entities? It will be related to the most common shape used to describe a core set of related entities and their essential relations. It may align strongly with a context graph around the internal entities. But different from an ontology, getting it right the first time is not critical. If one wishes to include another relationship at a later date, it can be added on top of/into the simplex to create a simplicial extension. It could arise from a rule-driven inference, or a local domain requirement. Since Knowledge Graphs are flexible, having a flexible semantic only empowers KGs more.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LlREbSkiu3sG2K7PHABbOw.jpeg" /><figcaption>Fig 2. Hulled 3-SImplex with no relation between Primary Care Physician and Patient slected OTC. For prescribed treatment, the PCP-prescription induces a sixth edge to be added, thereby creating a full 3-simplex.</figcaption></figure><p>Let’s assume there is once again the <em>doctor1-patient-disease </em>relation (with PCP), but now the patient buys an OTC treatment without the doctors knowledge inducing the relation, <em>patient-disease-OTC_treatment</em>. The 4 vertex relation is “hulled”, consisting of 1 diagnosis 3-relation, and an abutting self-intervention 3-relation. If a treatment is prescribed by the PCP, a new relation is added between PCP and Treatment, turing the hulled simplex into a full simplex. Each of the 4 3-relations has meaning: <em>patient-disease-prescription, PCP-disease-prescription, patient-PCP-disease, and patient-PCP-prescription</em>. Some are not that informative and can be left virtual, but they often are useful in analyses with aggregate counts (via <em>shapes</em>).</p><p>What does one do if some entities in the full set of entities do not have natural relationships between each other? These can be skipped if necessary to produce what is known as a <em>simplicial complex</em> which is a collection of <em>hulled-simplexes</em>, where the highest simplex order of entities is less than the total number of entities, which often is the case. Ontologically this means that there are shapes that do not have every entity neighbor relating to each other. As before, if one wishes to include a relationship that was left out by a data steward, it can be added on top of the complex to generate a natural simplex extension. <em>Shapes</em> in <em>SHACL</em> work easily with such complexes.</p><p>The indexing of either shapes or simplexes should be thought of as a convenience extension, that is some applications is even essential. Indexing all k-relations for <em>n,m,…,q</em> size of type entities could indeed create a combinatoric explosion, 𝑶<strong><em>(nm..q)</em></strong>, if all multiple association are allowed. In reality this will almost always be smaller (~n*2ᵏ⁺¹ ), even compared to QKV-attention embedding spaces for LLMs. More realistically, if one of the types in a <em>shape</em> is a leading key entity (e.g., patients and their medical associations), the number of <em>k-simplexes</em> will always be of order 𝑶<strong><em>(n)</em></strong>.</p><p>What about all the other <em>sub-shapes</em> and <em>(k-j)-simplex</em> relations inside the largest <em>k-simplex</em> (MCSS)? By following the simplex semantic I described above, they can all be inferred directly and generated on demand, since these sub-relations all follow the MCSS description and are unique in this context. For instance, <em>doctor-patient-disease</em> and <em>doctor-patient-treatment</em> fall naturally out of the <em>doctor-patient-disease-treatment</em> MCSS. The only thing needing to be added for improved semantics, the prepositions-predicates, such as <em>with</em> for <em>patient-with-disease</em>, or <em>treats</em> for <em>doctor-treats-patient</em>, or <em>for </em>in <em>doctor-treats-patient-with-treatment-for-disease</em>. This would be a fast look-up by entity type and substitution <em>if</em> one does not need to have all inner simplexes preemptively reified. The total number of n-way-relations in a <em>k-simplex</em> is: 2ᵏ⁺¹–1, which is exponential, so an auto-shape-typing, ensures semantics are efficient and reduce size explosion with labels. One can think of this <em>simplex </em>as a<em> local dynamic ontology</em>, having it’s own index.</p><h3>Summary</h3><p>Shapes and shape semantics (possibly simplex semantics) are a new methodology for rethinking knowledge graphs and their ontologies. They appear to have a connection with how LLMs input and respond to prompts, possibly giving us insight to why <em>attention</em> works so well. I’d like to suggest the entire reason we were told to engineer<em> prompts</em> and <em>utilize natural language with agents </em>(or any natural human language<em>)</em>, was a simple way to informally specify shapes that an LLM can digest to match how it organizes tokens and literature, as well as how it defines <em>attention</em>. These <em>shapes relations </em>cannot be fully pulled out since no label or id exists to define them, but they illustrate how a complex chunk of relations appears to work so well between GPT models and humans. I propose readers to look around in your fields of interest, and see if this can provide more useful insights into how we organize information and knowledge. Some examples I have worked on for a few different information projects include:</p><ul><li>Drug development workflow design and operational analyses</li><li>Assembly of multiple components into subsystems (e.g., engineering and construction projects)</li><li>Task directives for agents which include many sub-tasks</li><li>Biomolecules associated by hierarchical complexes (see Fig 3)</li><li>Medical procedures involving multiple phased procedures</li><li>Organization of mathematical proofs</li></ul><p>Regarding biological mechanisms as in bullet 4, a key feature of why life exists in the first place (think programs without processors), enzymes play a major role in every organism, and they modify chemical reactions so as to control them exquisitely. It begins with the Michaelis-Menton model of enzyme kinetics, which shows how enzymes are able to control chemical reactions via their binding sites and rate coefficients. A 2-simplex composed of the substrate, product, and enzyme cleanly organizes all the fundamental without need for excessive objects. Here the edges and faces are part of the equation:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/890/1*-1qq0E2DuYAJiJ5o8y2UDQ.png" /><figcaption>Fig 3. Michaelis-Menton enzyme reaction as a simplex of 3 molecular components: Substrate (S), Product (P), and Enzyme (E), where |E| determines Vmax, and SE association determines Km. The intermediate complexes formed (SE, PE) are edges, the final reaction rule is the edge on top, and the entire enzyme-limited catalysis defining the equation is the 3-face.</figcaption></figure><p>Many more examples exist from many diverse areas: writing of screen plays, complex recipes (other than pizza mixins), analysis of written works, comprehensive examination of evolution over time, image recognition (ala TDA), and many more. Linearization of models (graphs) may be useful for checklists and chronological steps to follow, but are not a good source of knowledge that includes inter-dependencies.</p><p>What is the risk of not fully comprehending the implications and benefits of shapes and simplex semantics along with our expanding intelligence and knowledge systems? There an obvious few things, that look small at a first, but could eventually become substantial:</p><ol><li>Ontologies continue to be misunderstood and the “ontology wars” continue while alternative solution grow (<em>which uber-ontology should we use?</em>).</li><li>Extremely large indexed databases that become more difficult to search by 2,3,…-way associations, including why LLMs are growing at a cost of longer training cycles. I believe much of the hyped expansion of data centers is because nobody really understands the internal representation issues of data and facts. Billions could be saved by using data structures more effectively.</li><li>Semantic modeling by pairwise associations in KGs becomes unmanageable, and everyone gives in to using super massive language models (SMLM with over-priced<em> pay-as-you-use</em> services). Society could be paying for inefficiencies it could avoid, because the operators of AI do not see any advantage for streamlining AI if capital is available.</li><li>LLMs will evolve separately of other data and structured resources, costing $billions many industries (like pharma) as they continue to invest in AI without knowing where it best can be applied or why it works / doesn’t work for some things.</li></ol><p>What started as a set of principle for better design and use of knowledge representations, may become linked to the larger issues around future of affordable and more environmentally-friendly intelligence systems. If the mantra that the <em>future is AI</em> is true, then surely anything that can go beyond the current model will be worth even more. Most acknowledge we do not have a full grip of what (artificial) intelligence is and where it will take us. It might allow us to avoid some bad techno-dominant consequences some experts have been warning about. We need to examine the current strategies for computational intelligence based on old models, and determine if there aren’t better opportunities near us.</p><h3>Footnotes</h3><ol><li>It is easy to show that the real value of today’s ontologies is its ability to specify the relations a (conceptual) entity have to other entities (object properties, edges) as wells to data (data properties). A <em>Concept</em> without any specific relations is equivalent to a empty node. So it stands to reason that the <em>Concept</em> labels and taxonomic hierarchy (subsumption) is a “secondary” definitions actually specified by the relations it has. In other words <em>Shapes define (labelled) concepts, not the other way around.</em></li></ol><p><em>Editor’s Note: This article was updated on July 21, 2026, to include additional references for further reading, that describes an analysis of Yoneda and formal ontologies.</em> [4]</p><h3>References</h3><ol><li>R McEntire, P Karp, N Abernethy, D Benton, G Helt, M DeJongh, R Kent, A Kosky, S Lewis, D Hodnett, E Neumann, F Olken, D Pathak, P Tarczy-Hornoch, L Toldo, T Topaloglou. <em>An Evaluation of Ontology Exchange Languages for Bioinformatics</em>. Proceedings of the Twentieth International Conference on Machine Learning, 2000. ISMB AAAI.</li><li>Chazal F and Michel B (2021) An Introduction to Topological Data Analysis: Fundamental and Practical Aspects for Data Scientists. Front. Artif. Intell. 4:667963. doi: 10.3389/frai.2021.667963</li><li>A Oudshoorn, M Ortiz, M Šimkus, <em>SHACL validation in the presence of ontologies: Semantics and rewriting techniques</em>, Artificial Intelligence, Volume 352, 2026. <a href="https://doi.org/10.1016/j.artint.2026.104483">https://doi.org/10.1016/j.artint.2026.104483</a>.</li><li>Seremeti L, Kougias I, <em>Yoneda Philosophy in Engineering</em>, <em>International Journal of Engineering Mathematics</em>, 2013, 758729, 11 pages, 2013. <a href="https://doi.org/10.1155/2013/758729">https://doi.org/10.1155/2013/758729</a></li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=397162641eea" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>