According to the Oxford American College Dictionary, a principle is “a general scientific […] law that has
numerous special applications across a wide field,” describing “a fundamental quality determining the nature of something.” In its most general definition, a principle captures relevant and simple regularities.
Ever since we founded SCIoI, we have been conducting research in the fields of both natural and synthetic intelligence. Within the synthetic (i.e. artificial) approach to intelligence research, we were able to identify recurring patterns, which represent possible principles of intelligence. In order to capture and test their knowledge about a phenomenon, scientists use models. Decriptive models capture the measurable properties of a phenomenon, while mechanistic models are descriptive models that also capture the process that brings about the phenomenon. The recurring patterns we were able to propose are mechanistic principles, “general scientific […] laws that have numerous special applications across a wide field.” These laws don’t only define the intelligence of a system, they also characterize properties for the construction of potential artificial agents. For each of these proposed principles, we were able gather empirical support across species, tasks, environments, and levels of abstraction, proving that these principles hold true under many aspects. Our research convincingly shows that mechanistic principles are a suitable approach to understand, characterize, and synthesize intelligence, and in a very general sense, they give us an idea of how the components of intelligence, i.e. the different disciplines, interact with one another.
In order to be considered a principle, a scientific law needs to possess certain properties, which are listed here.
Based on these properties, we are proposing eight recurring patterns, or principles, of intelligence (and a ninth one is on its way):
These candidate principles are meant to bring us closer to defining aspects of intelligence, and it is important to note that they don’t all have to be simultaneously part of a system for it to be considered intelligent. Some of the proposed principles (or “recurring patterns”,) such as the principle of active interconnections, seem to be present in all intelligent systems, while others do not necessarily need to be. It is really not all that surprising. Even literature (think: myths, tragedies, archetypes of heroes etc. that keep recurring), cooking recipes (think: roux turning into millions of different sauces by instantiation of a template), and music (think of stilmas, minimal units of music that help us define a musical style) work that way.
Why do we need principles?
Principles guide the formulation of mechanistic hypotheses about mechanisms of behavior generation in specific instances, i.e., species, tasks, cultures, developmental stages etc.
Their purpose is to reduce the space of hypotheses we must investigate. They provide conveniently encoded prior knowledge so we can produce hypotheses that are already supported by prior evidence.
Think of the principles as the periodic table of elements for intelligence research. We will continuously add new principles to the table. All of intelligence research will either produce new principles or directly find mechanistic explanations of intelligent behavior by leveraging the known principles to formulate hypotheses that are highly plausible based on prior evidence.
This substantially accelerates intelligence research because the existing principles guide us towards the space of plausible mechanistic explanations.
Natural and synthetic intelligence –– vs. no intelligence at all
When we think of natural intelligence, we immediately think of the animal brain. Humans, dolphins, cockatoos, penguins, but also mosquitos and worms are just some of what we consider intelligent beings, naturally possessing a nervous system that allows them to satisfy the principles described above. Synthetic intelligent beings, such as robots or learning models are based of natural intelligence, and their behavioral cognitive bases are comparable to the ones of natural systems. But intelligence manifests itself in other forms too. Think of a school of fish or a flock of birds, or even a swarm of robots. Here, the interactions between individuals form a sort of collective brain that transcends from the individual itself and guides group decision that are sometimes more accurate than individual ones. So, when can we say that a system is not intelligent? It is not intelligent when it does not satisfy any of the principles above: Examples include engineering systems such as cars, calculators, surveillance cameras, or thermostats, which follow mechanical rules without however possessing the adaptability of what we call intelligent.
Example of a principle (in principle)
The notion of recurring patterns abounds in mechanical engineering. These recurring patterns occur at different levels of abstraction. To describe a mechanical system, we can talk about a motor, a gear box, and an output. This in itself is a recurring pattern. It is not a fully specified system but it tells an engineer how to construct a generic system that takes an energy source and converts that energy into a process towards a goal (think of the steam engine, for example). We can now go one level deeper into the concept of “gear box” to find more recurring patterns.
There are many different types of gears: worm gear, helical gear, bevel gear, spur gear, rack and pinion. Each of these types represent a recurring pattern. We can use these patterns to build specific gear boxes. This allows us to leverage existing knowledge rather than having to solve each problem from scratch before. It works well, because the different types of gearboxes broadly cover a wide range of application cases.
In the Figure below, we hypothesize three different mechanistic structures that produce behavior in mice, fish swarms, and robots, respectively. These mechanistic structures are illustrated using the visual metaphor of spur gears (see rectangular boxes), which are a type of cylindrical gear with straight teeth that are cut parallel to the axis of rotation. They are the simplest and most common type of element in mechanical drive systems. You can see that the specific instances for each species differ, for example, in the size of the gears, orientation of the gears, whether the gear teeth are straight or slanted, what material they are made out of, how many teeth they have. Nevertheless, they follow some kind of blueprint that captures all possible instantiations of spur gears (the rounded box contains the template for all spur gears with many parameters to be instantiated in any specific gear system). We call the blueprint “principle” and the specific mechanisms implemented in the species “instantiation of the principle.”
The principle describes an abstraction of the more specific and more concrete instantiations that can be observed in any specific system and situation. The black arrow means “X is an abstraction of Y”, where X is the beginning and Y is the end of the arrow.
.
The principle of “gears” offers a templated solution for changing the direction and speed of a rotation. It does so by using two connected gears. But all the details of how the gears should be structured and arranged are left out. The specific direction and specific changes of speed will be different for the concrete instances. To obtain a specific mechanism for generating behavior, we therefore must instantiate the “gear template” (or principle) by specifying the diameter of the gears, number of teeth, relative arrangement, material, and so on.
A principle of intelligence, by definition, must apply across some range of species, tasks, scales (individual, social, collective), developmental stages, etc. — in short: across some points within the space spanned by the dimensions of SCIoI’s comparative approach.
The easiest way to find principles of intelligence
Since principles – by construction – occur repeatedly across species, across levels of abstraction within the organization of an individual, across scales (individual, social, collective behavior) and contribute to many behaviors at different developmental stages, we should be able to observe their consequences in many different situations.
The easiest way to find hypotheses for principles of intelligence is thus in comparative studies across all the different dimensions listed in the previous paragraphs. And the more diverse the comparison is, the easier it should be to identify commonalities. These commonalities can quickly lead to hypotheses about principles. This is like a key to unlock the secrets of intelligence. The more scenarios we compare, the more likely it will be that something that is shared by all of these scenarios contains some important contributor to the generation of intelligent behavior.
Please note how fundamentally different this is from the traditional disciplines of intelligence research. Psychology, for example, studies only humans, making it difficult to identify commonalities or recurring patterns. The answer in psychology has been to study narrower and narrower questions without a reasonable plan to merge the accumulated details into a coherent whole.
Or take neuroscience: Here a single level of abstraction, probably the lowest level of abstraction possible, is studied. Confining oneself to this low level of abstraction makes it exceedingly difficult to find patterns in the first place, as they will be intermingled with many detailed phenomena revolving around idiosyncrasies of the neural substrate.
Even in behavioral biology, most researchers are focusing on specific species and even on specific behaviors, again possibly obfuscating the recurring patterns we call principles that would shed light on the mechanisms underlying intelligence.
