Avoiding the Ontology Trap: How biotech shows us how to link knowledge spaces
Unsupervised Learning for Humans, from bottom-up networks to namespaces
How we group things together (or separate them) says a lot about how we think about problems. What we choose to see as similar or different is highly contextual, if not merely spurious, but when we try to make sense of something we are trying to justify those decisions. We sometimes refer to “namespaces” in IT to refer to independent “chapters” of a system, where different rules apply. In databases and programming, we create “tables” or “classes” of data, with a pretence of universal meaning. In books, we classify ideas by subject category, by fact or fiction, etc. But in knowledge graphs, we throw away a lot of those classifications and let everything float in a single space of ideas. There is both freedom and chaos in this that allows lateral thinking to join distinct areas. Topic maps (an early form of knowledge graph) tried to combine both classification and association to capture the best of both worlds, but in doing so marginalized itself both by requiring too much expertise to model, and by falling foul of the ontology trap. In SSTorytime, we have two mechanisms for defining conceptual spaces: chapters (like species) and context lists (like genetic traits).’
The SSTorytime project is a small act of defiance against those who sponsor this great subjugation of human agency: a tool to encourage a methodology for digital knowledge representation aimed at human development with power tools. Please join in!
Avoiding the junk heap
As the sum of human knowledge mounts, the more it resembles a junk heap of things we are simply unfamiliar with. The slogan “it isn’t knowledge if you don’t know it” expresses the idea that you need to know that something exists, and then you need to know it well enough to know that you need to know it better! It’s the circularity of reinforcement learning.
To become truly “knowledgeable” in any subject, requires work and dedication–but even knowing enough to look something up, or ask a question requires us to lift ourselves by our bootstraps and know something about it to begin with. The problem is that we have limited cognitive faculties, so we have to measure knowledge in manageable chunks.
Today, technology seduces us to forgo that setting of boundaries, to ignore the husbandry of knowledge, to give up our personal agency and subordinate ourselves to limitless vistas of information that we can’t hope to digest. To use such information responsibly without a basic knowledge already, we need to at least build a basic familiarity with it, to map out its main ideas. “AI” can paper over some cracks, but it cannot and will not solve the crisis we face over learning, as a society, nor guarantee our future understanding. Knowledge cannot work like Jack In The Box.
I’ve written a lot about the nuts and bolts of knowledge representation over the past months and years–about how to combine language with graphs to express ideas. In this essay, I want to think about how we partition knowledge or separate concerns and package them for consumption. The basic answer is to maintain a relatively hard boundary for different “chapters” that seem self-contained, and more of a soft boundary for different “contexts” that belong to those. But associations can always tunnel through these barriers to reveal the unexpected.
Neuroscience and the Dunbar limits
In 2022, I boldly announced to a room of network and neuroscientists that the human immune system is the second most advanced reasoning system in the mammalian species. This was not a particularly controversial statement for a room of mainly biologists and mathematical physicists (though it was probably unexpected), but for most people it probably is. That seems like a controversial gauntlet to start with, and yet understanding a little biology turns out to be the key. We’ve narrowly come to associate thinking only with our own human minds, making ourselves unique and special in a dangerously ignorant way. Yet, long before humans walked the Earth, life had taken billions of decisions without so much as a brain to stand on!
The vertebrate immune system is, in fact, a highly sophisticated and distributed system: a diverse battery of cell types forming a vast ecology of processes that is deviously complex. It isn’t a solid lump of neural connections–it exists in a largely fluid state. It is a simple instance of what people are calling “agentic AI” today. Collectively, it has memory and enough discriminating chops to identify dangerous pathogens, then manufacture antidotes (even for substances that have never previously existed in the history of the planet), and bring us back to a state of “health”. However, in spite of this apparent expertise, no single cell has any intelligence nor even a notion of what it’s doing. It works, as the entire biosphere works, by the smallest of language models: recognition of patterns, supporting complex ecological interactions that seek out stable equilibria. It’s as if voting in an election could manage our health. And yet this is how knowledge works too.
We "vote" for our knowledge by repeatedly thinking about it, or by working with it actively. Little do we realize, most of our memory is actually outside of our brains–in the things we build, communicate verbally, and write down. We leave stigmergic trails, just like ants. By contrast with brain reasoning, our broader societal computations, such as bureaucratic processes, which involve the work discipline of filling forms and paper trails (a particular pain in the upper and lower spinal cord) — are also reasoning processes. They may or may not involve human brains (the jury is still out on that), but the process performs reasoning nonetheless.
To manage these knowledge spaces, we need to understand all these processes–but it costs us. For every issue we have to deal with, we have a trust/mistrust instinct that acts to limit the cost of our rather expensive cognition. The so-called Dunbar numbers come into play, as representations of our cognitive limitations (see my recent collaboration with Robin Dunbar). Any memory structure or knowledge management scheme has to be compatible with these limitations, else it will simply pass us by, flying high over our heads onto the junk heap.
What’s interesting about the immune system, for such a simple collection of cells, is that it has no obvious modularity or separation of concerns on the surface. It has older non-specific parts and more recent targeted upgrades, but these coexist in a muddle and only occasionally step on one another’s toes. In fact, on closer inspection, there are modules like the skin and the bloodstream, search functions and neutralizing weaponry, but they all use overlapping technology based on a few principles. In certain areas, the immune system is highly precise, but in others it is quite elastic and tolerant of muddle. What can we learn from this?
Each of us has our own view of things and our own needs. We don’t manage information like a gigantic central brain. There is a reason why we are individuals with different brains.
The function of a knowledge graph, in such a process, is not to act as a database of memories, but to encapsulate many fragmentary episodic experiences in such a way as to allow us to make cross connections when they might occur. Yet, our familiar strategies for knowledge management are based on a taxonomic separation of concerns and the subdivision into authoritative hierarchies.
Taxonomy: thinking inside the box
“Among insects there are innumerable instances, thus Linnaeus, misled by external appearances, actually classed an homopterous insect as a moth.”
–Darwin, Origin of Species
We see shapes in clouds, faces on cars. It’s a symptom of what makes us smart. Language is made up of such patterns, and we represent meanings by stacking the patterns in different ways. One mistake we make is to pretend the world is static, invariant, or constant, and try to sort it into boxes.
The two culprits of classification are taxonomy and ontology. They are ultimately self-defeating attempts to regulate and legalize the boxes of our thought, and offer short term relief from the chaos of having too much to know–by basically fencing off the freedom to explore analogies and associations that give knowledge depth.
Taxonomies became trendy when Swedish biologist Linnaeus introduced in the biological realm (see my book In Search of Certainty). Taxonomies are not unique. In terms of Graph Theory, they are just approximate spanning trees: i.e. branching processes that form a map of a connected web of things, starting form a single root node, which is arbitrarily chosen.
There is no unique semantic spanning tree for an arbitrary knowledge graph, except for trivial ones–no universal truth about the division of fauna and flora into species purely on the basis of their external attributes. Similarly, there is no universal set of concepts to subdivide knowledge management. These are merely ad hoc ways of spanning a collection of connected ideas, chosen by convention or happenstance, and often even imposed onto others as a kind of attack of authority. We associate hierarchical organization with authority, even when it is fictitious.
When we write books, we use two methods for mapping out information: the Table of Contents or chapters, which is a systematic summary of themes (assuming that chapters are designed to keep related issues close together), and the Index–which is a random access lookup table, like a telephone directory for search terms. Search engines are automated indices. Taxonomies attempt to mimic the behaviour of a table of contents. Ontologies are just glorified taxonomies with a few rules about “see also”.
The great conceit of a branching hierarchy representation of knowledge is that the world is a simple unfolding of ordered states. But we know that it isn’t. Whenever we come up with a rule that fits some exterior characteristic, it turns out to have an exception, because it is merely a spurious correlation–not a causal explanation. The duck-billed platypus is the classic example: a warm blooded mammal that lays eggs instead of giving birth to live young. When rumours of the Platypus first reached Europe, it was thought to be a hoax, because the model of species had acquired a certain authority. This is why graphs allow us to make more meaningful maps, adding details about what functional properties a space has. We’ve seen this idea in Google Maps: tell me where the nearest MacDonalds is, or where I can get new shoes. Knowledge maps apply the same thinking to any abstract spacetime or possible process.
Horses for courses, namespaces for individuals
To support individual knowledge, and therefore individual agency, every one of us needs our own prosthetic knowledge assistant based on our own lives–not one that takes away our experience or thoughts for us and replaces them with a generic imposter, but one that we feed with our own thinking and that we can search later to help recall our previous trains of thought. For a social group of people, the structure of a knowledge ecosystem might look something like the image below.
We can imagine every individual, person, user, up to and including collective organizations equipped with their own individually curated knowledge resources (this is the goal of the SSTorytime project). Each will be most useful to the person(s) who creates, interacts with, and uses it. It can be shared with others for inspiration, but someone else’s vision is never as powerful as our own.
The separation of concerns here is to prevent one person’s recognition from interfering with another person’s recognition–while still allowing knowledge to be shared. It sounds easy enough, but how can it be achieved without everyone sharing a common language–a so-called common ontology in the philosophical sense?
We know the answer to this already, from a totally different kind of informatics: bioinformatics. Our modern knowledge of the language of DNA has opened our eyes to the way that knowledge actually works–by exposing the structure of an interior language based on process causality from semantic building blocks.
DNA and mapping semantic spaces
We can roughly divide the active elements of our experiences into three kinds of “happenings”: events, things, and concepts (forming the convenient mnemonic “etc”). See Designing Nodes and Arrows for Semantic Spacetime. These three entities are the building blocks of processes, which is what DNA represents within the context of a cellular apparatus. The much touted genetic code, written in DNA strings, is an alphabet for shaping a process, in the same way that knowledge graphs are about processes consisting of events, embellished with facts about things and their functional attributes (concepts).
In SSTorytime, we have fragments of concepts that label clusters of relations, e.g. see the figure below:
The self-contained or “hard” barrier is the chapter “neuroscience brain”. The soft context refers to the user’s current thinking: “brain waves” or “oscillations”. A comma separated list acts like a local index for connecting random associations with a group of topics: alpha waves, etc. But the boundary isn’t absolute. Any one of the terms can also exist in another chapter, e.g. gamma waves might occur in a chapter about physics of radiation. The key is a loose connection between apparently random compositional fragments and the meaning of the whole in context.
This method of encoding persistently functional information, by composing small fragments, was rediscovered in linguistics by the Greeks when writing in their alphabeta. Indeed, one can say that biology is an application of linguistics to chemistry. The Ancient Greeks knew the power of writing as a representation of knowledge, and further realized that–although the Egyptian and Chinese glyphs were pictorally attractive and associatively meaningful (at least as long as the imagery was familiar)–it was more efficient, even pragmatically simpler, to build a system based on fragments that were easy to copy. From there, one could construct any meaning by combination and association. And so, with the fervor of a particle physicist who gleefully smashes atoms to discover its “smaller parts”, letters and phonemes, they could extend, reconstruct, and analyse meanings in terms of those genetic signatures.
But that’s only half the story. Fragments are not just a way of making a catalogue of parts to order. The recombinant soup of those fragments forms an enveloping “context” around the central elements of a story — it is far from being random. It isn’t simple, but it is also not random. In fact, it becomes a key to unlocking the interpretation of each present “moment” in a scene. Just like the :: …context … :: tags in SSTorytime.
Automated inheritance with fluid semantic spaces
If mating is a way for groups of genes (as partial ideas) to survive by mixing and selecting, then we can make ideas survive (as knowledge) by enabling a user to find ideas through their genetic components too. The ultimate type of chapter will allow selection to hone relevance more sharply.
Our comprehension of ideas, in a knowledge representation scenario, is like a scene in which key players are floating in a primordial soup of fragments that give the meat and potatoes of the story flavour (apologies to vegans, and their poetry, who might only understand half of that argument). We can see this in action, with the automated fractionation of text in the SSTorytime project tools. This quote from Darwin’s Origin of Species shows a few of the large list of fragments.
The relationship between the nearby fragments and the full story is what the knowledge graph can represent–and the Semantic Spacetime links can measure. The full story is analogous to the full description of a species as an idea (in relation to its process for propagation). The fragments represent universal access points that are free of that binding context.
We should not confuse the significance of the fragments with the significance of the whole, any more than we should confuse a member of a species with its individual genes. They are not equivalent. To unlock paths through a network, the keys are not small items, like the keys to doorways. It’s more like solving a maze than opening a door. There are several ways to do it, but one way (used by ants and other organisms) is to maintain an actual living population of active agents spanning the problem to create a trail. It’s a physical representation of a story. This is why knowledge cannot be a dead thing.
Effective languages at the right scale
Defining namespaces is about deciding where the boundary between different "scales" of attention should be placed to lessen the cognitive burden.
In computer science, the issue of scale in process fragments has led to a wide range of languages for the same idea, in different processes. If general purpose programming languages (often called “Turing Complete” languages these days) are the Lego sets of process automation, then the Domain Specific Languages (e.g. SQL, CFEngine, etc) are an attempt to optimize a higher level, specifically approximated digital twins of certain selective processes and their data sets, by identifying their higher level patterns.
In physics, one has so-called effective theories that are suitably idealized approximations of phenomena without undue complication. The way we create these idealizations is to break processes up into their atomic pieces, like a DNA analysis. Domain Specific Languages are the “effective theories” of computer processes. Like all effective theories, they eventually prove ineffective in cases where we need to dig to the very roots of a problem–but until then, they expose the essential “semantics” of phenomena through stable patterns. They are pattern languages.
For physics, at each “scale” (usually characterized by some typical distance in space, interval of time, or rate of change) there are new phenomena to describe and explain. At the bottom of the hierarchy (which we assume exists) there are “atomic or indivisible things”. We have the table of elements, the chemistry of molecules, from which we get the science of materials and the biology of cells etc. Depending on your point of view (which is the key point) these are either fundamental or merely derivative phenomena. In practical terms, each new scale introduces new information and therefore forms an independent “cell” in a loose hierarchy of scales for knowledge.
At each scale, it’s usually possible to identify “agents” or components that play certain roles in phenomena. This was the insight that led to Promise Theory. However, like the story of the platypus, we shouldn’t expect the world to form neatly partitioned non-overlapping nation states. What one finds in one phenomena invariably has analogues in others.
Postscript: From AI Oracle To Eloi
The idea of having a kind of oracle containing the sum of human knowledge so that humans could be its children has been parodied in science fiction many times. Would it help us to really know anything, or would we simply worship this fact toaster? If we could all be Wikipedia, or could all share the same brain, then there wouldn’t even be any need to talk about stuff anymore, we would all know exactly the same thing–instant telepathic ascendance. But then there would also be no one to challenge us and our knowledge would be dead.
To keep knowledge alive, it has to be the very lifeblood that pumps through every individual’s cognitive veins and keeps society alive. In a knowledge society, there can be no “developed” countries–only continuously developing and striving frontiers.
Knowledge management might not quite deserve the status that has been thrust upon it in recent times by hungry financial sponsors of “AI”, but it is surely one of the most important topics for humanity to address. The information age has made that easier than at any time in history to forgo our own personal development, so we need to be ever more vigilant in approaching knowledge and how we use it. By embracing our subordination to dead archives, we have already begun on a path, stepping closer today to a new “Dark Age” than at any time since The Enlightenment.
To see how it works, you can try browsing SSTorytime and the examples included in the distribution.
