Two notions of medium and their implications for intelligence

In biology, the term medium is defined as a substance that a biological system exists or grows in or that it travels through. In physics, the term medium is defined as a substance that propagates and transmits the energy from one location to another. These two notions of medium in biology and physics are distinct, yet in reality, their referents frequently coincide in the same material systems in the environment. The air, a medium for terrestrial animals in the sense defined in biology, is a medium that transmits light, mechanical waves, and diffuses molecules from the source in the sense defined in physics. A medium that surrounds each cell in the living body in the sense of biology, at the same time, is an excitable medium that propagates electrical events, mechanical stresses, and the variety of chemical molecules in the sense of physics. In this article, I discuss the implications of this coincidence of the two notions of medium in the real world for the evolution and development of intelligent systems.


Two notions of medium
With intelligent agents, altering the conditions changes the activity displayed indefinitely, but not the end reached.The defining characteristics of intelligence is, as James [2] famously put it, "the fixed end, the varying means."What are the ecological resources available that could potentially provide opportunities for the evolution of intelligence in this James' sense?This question may sound superfluous, for what is responsible for intelligence is usually attributed to internal structure of the system that exhibit intelligent activity.Yet, as Robert Rosen [3] stated succinctly, a function of an organism can never be understood in terms of its internal structure simply because "a function requires an external context; a structure does not."The functional act of pursuance of ends and the choice of means for their attainment is partly dependent upon external factors beyond the control of the actor.For an intelligent system such as a living organism, whereas being an intricate mechanism in itself, is one whose movements is tailored to things in the environment outside of itself, much as fish in the pond; and these external things 1292 (2023) 012022 IOP Publishing doi:10.1088/1757-899X/1292/1/012022 2 are as much constituents of the process of activity as is the living organism itself [4].In what follows, among others, I shall focus on one class of material reality with which the development of activity of biological systems is inextricably linked, namely, the medium of the environment.
The term medium means different things in biology and physics.In biology, the term medium is defined as a substance that a biological system exists or grows in or that it travels through.The air, for example, is a medium for terrestrial animals, and the water is a medium for aquatic animals.A medium permits respiration, ceaseless chemical exchange of substance that sustains life.At smaller scales, for cells that collectively constitute the body of an organism, the extracellular matrix and the extracellular fluid can be considered as a medium.A medium is a substance that surrounds each structural and functional unit of biological systems at its scale.
In physics, the term medium is defined as a substance that propagates the energy from one location to another location.Various forms of energy are propagated in different media.Sound, for example, is a compression mechanical wave that is propagated through a medium such as the air, water, and various elastic solid materials.Unlike mechanical waves, electromagnetic waves such as light may occur in a vacuum as well as in a material medium such as air and water.In physics, a medium is where different structures arise from various physical events, while its substrate material being unchanged.For example, you can listen to music and then have a conversation with a friend in a room, without exchanging the air in the room for each event.
These two notions of medium in biology and physics are distinct, each of which has its own emphasis that captures different aspects of the world, although they are not mutually contradictory.The point I would like to draw attention is the fact that, in reality, the referents of these two senses of medium frequently coincide in the same material systems in the environment.The air, a medium for terrestrial animals in the sense defined in biology, is a medium that propagates and transmits light, mechanical waves, and diffuses molecules from the source in the sense defined in physics.A medium that surrounds each cell in the living body in the sense of biology is, at the same time, an excitable medium that propagates electrical events, mechanical stresses, and the variety of chemical molecules in the sense of physics.
What are the implications of this overlap of two senses of medium-biological and physical-for intelligence?What kind of opportunities can be potentially made available in this juncture of the two notions of medium in the real world?In the present paper, I shall discuss how an environmental medium, under certain conditions, could carry information about things and events in the world that can be potentially detected by the perceptual system of an agent.I shall also point out the parallel between the opportunities provided by the physical overlap of these two notions of medium, and the development of so-called physical reservoir computing, a framework that exploits the complex dynamics of physical systems as a computational resource [5,6].

Liquid medium
The water is a medium for aquatic animals in the biological sense, which the animals grow in and travel through.At the same time, the water is a medium in the sense of physics where various patterns of water movements-from ripples and eddies to tidal waves-take place.Through evolution, most aquatic animals living in water have developed sensory systems to discriminate flow patterns in their aquatic medium [7].These flow patterns can be caused by different source events, including the animal's own swimming movements, inanimate objects, conspecifics, predators, preys, and weather events.For example, the water movements that arise from faststarts in teleost fish consist of multiple jets, based on which aquatic predators can detect not only the presence of a fish and its size, but also the direction in which it escaped [8].Harbor seals use their vibrissae to haptically discriminate the flow structure left behind by predator or prey, and detect the size, shape, and motion path of the object that gave rise to the structure 3 [7].The water movements caused by the swimming motion of different species of fish have distinct flow signatures which can remain for several minutes in the water (Figure 1).These flow structures provide piscivorous predators at a distance with valuable sources of information about the presence of a fish of suitable size as well as the swimming style or the species of the fish that have passed by in the recent past [7].Unlike vision, hydrodynamic perceptual systems in the aquatic medium permits sensing the past event without the aid of internal memory, because perturbations in the aquatic medium, which gradually fade over time, can specify the events that occurred at an earlier point in time.
The flow patterns formed in the aquatic medium provide the aquatic animals with the openended opportunities for selectively attending to their different features to detect various events that are relevant to the habits of life of the animals.When we consider those two aspects of medium-that which surrounds organisms and that which transmits energy-hydrodynamic perception seems almost a natural consequence of their overlap in the aquatic medium.Among the conditions that potentially make hydrodynamic perception possible are as follows [9]: (1) disturbances of pattern in the aquatic medium are sensitive enough to separate different source events, and (2) the perceptual system of an aquatic agent has an enough resolution to discriminate the different patterns formed in their medium.

Liquid state machine and reservoir computing
Independently from research on hydrodynamic perceptual systems, a computational neuroscientist Maass and his colleagues have arrived at some similar viewpoints: The perturbed states of the excitable medium could potentially make available the reservoir of information that can be exploited by intelligent systems to perform various tasks.Maass et al. [10] identified that the key challenge to understanding intelligence is to find the adequate conceptual framework that can account for how agents detect in real-time the equivalent state from the continuously changing transient stimulation which may never repeat.In tackling this challenge, Maass took seriously the material environment of neurons-the fact that "the neurons in our brain are embedded into an artificial sea-environment, the salty aqueous extracellular fluid which surrounds the neurons in our brain" [11].Using the metaphor of liquid, Maass et al. [10] wrote as follows: "consider a series of transient perturbations caused in an excitable medium . . ., for example, a liquid, by a sequence of external disturbances (inputs) such as wind, sound, or sequences of pebbles dropped into the liquid. . . . the perturbed state of the liquid, at any moment in time, represents present as well as past inputs, potentially providing the information needed for an analysis of various dynamic aspects of the environment.In order for such a liquid to serve as a source of salient information about present and past stimuli without relying on stable states, the perturbations must be sensitive to saliently different inputs" [10].
Fundamental to Maass and colleagues' idea is the overlap of the two notions of medium: The medium of the neurons in our brain-the extracellular fluid-in the sense of biology is at the same time an excitable medium whose state can be perturbed by various external disturbances in the sense of physics.Using perturbations caused in a medium as informational resources, Maass et al. [10] proposed a computational model called the Liquid State Machine, which maps in real time some function of time (e.g., an aspect of the environment that changes continuously) onto other functions of time (e.g., continuous adjustments of action in relation to the changing aspect of the environment).
In terms of architecture, a liquid state machine M consists of the following two components: The first component is an excitable medium called a liquid filter L M which generates, at every time t, perturbed 'liquid state' x M (t) in response to a preceding sequence of disturbances u(•).A liquid filter has the time-continuous property of fading memory whose state depends on the disturbances from some finite time window into the past, just as the perturbed states of the sea water lasts for some finite amount of time after a fish has passed by.A liquid filter needs not be customized for a specific task, just as sea water is not customized for each of the tasks that aquatic animals perform.The second component of a liquid state machine M is a readout map f M that transforms, at every time t, the current liquid state x M (t) (which expresses the past as well as the ongoing events because of the fading memory property of the liquid filter) into taskspecific output y(t).Stated another way, a readout map f M extracts the task-relevant features in the medium, and continuously adjusts the output in relation to the task goal.Depending on the task, the readout can generate an invariant response despite the fact that the high-dimensional dynamics of the liquid filter may never revisit the same state.Liquid state machines define the task-relevant equivalence classes for the dynamic liquid states [10], which instantiates one concrete implementation of intelligence in James' sense.
The conceptual framework of a liquid state machine was unified with so-called Echo State Networks-a machine learning approach independently developed by Jaeger that shares the fundamental operating principle [12]-under the umbrella label of reservoir computing.In the framework of reservoir computing, if the reservoir that exists independent of the task-specific readouts has rich and diverse enough dynamics to separate the different sources of disturbances, could make available the opportunities for real-time, task-specific control of the medium-readout system.Here, the role of a reservoir is to pick up the source event and propagating it into the high-dimensional space which could potentially render relevant features from the event more easily separable [13].
A medium-readout system can have the power for real-time computing on perturbations regardless of specific implementation or structure, if the following properties are satisfied: the separation property, approximation property [10], and echo state property [12].The separation property addresses the amount of separation between different disturbances of structure in a medium at later time points that are caused by preceding different sequences of events (e.g., different wake signatures resulting from different sequences of source events).The approximation property refers to the property of readout function that maps the separated states into the output that meets the functional requirements of the task.The echo state property is the property that requires the reservoir states to be expressed as a function of the previous input sequence only [5].
It would be worth noting that even though the task-relevant states of the reservoir are there to be extracted, they do not determine the readout response of the system.Just as the seal running away from the white shark may not attend to a pattern of water movement that specify the presence of salmon that can be preyed upon, the liquid states that are potentially informative about some events exist independent of its use.This allows for flexibility to select the means with which to attain an end, as well as to have multiple ends at the same time.In reservoir computing, it is possible to add multiple readouts to a single reservoir so that each readout extracts different task-specific information from the rich reservoir states in such a way to support different, multiple real-time controls in parallel.

Body as a medium
Turvey and Fonseca [14] hypothesized that hierarchical networks of prestressed tensile elements of the body that span from the macroscale to the microscale-from muscles, tendons, and other connective tissues to various micro-elastic structures such as a network of collagen fibers-constitute the medium of the haptic sense organs of animals.Like the air being the medium for sound, odor, and reverberating flux of light, the presence of isometric tension distributed throughout all levels of interconnected, multiscale networks of the body propagate various mechanical disturbances.Thereby, the opportunities arise for an active agent to spontaneously perturb the tensile states of hierarchical networks in such a way to separate the invariant patterns that specify the source of mechanical disturbances [9].
There is evidence that provides support for this scenario of active touch that takes into account the medium for the haptic perceptual system.In a study of children learning to read braille, Nonaka, Ito, and Stoffregen [15] followed the development of scanning movements of the reading finger of blind school children over the course of one year.Contrary to the previous assumption that braille reading is optimized when the scanning finger moved at constant speed, they found that the strength of long-range power-law temporal correlations in the velocity fluctuations of lateral scanning movement were functionally related to braille reading performance and its development (Figure 2).Their result suggests that, instead of degrading perceptual sensitivity, a certain structure of velocity fluctuations contributes to separating off the invariant information about the adjacent environment over the transformations of tissue deformation patterns, and that variations in movement of the scanning finger might be controlled so as to efficiently separate informative structures that specify the braille text.
In another example, Sornkarn and Nanayakkara [16] recorded the muscle co-contraction activity of the finger using the electromyography sensors when human participants were asked to palpate a soft tissue to locate a hard nodule embedded within it.They demonstrated that human participants varied the muscle co-contraction level (muscle tone) during manual palpation to estimate the depth of a hard nodule in the soft tissue.Furthermore, in the subsequent experiment using a soft robot, the team revealed that haptic information gain about the depth of the hard nodule can be enhanced by varying the internal stiffness of the soft probe of the robot [17].Their findings provided empirical support for the idea that the discriminatory sensitivity of the haptic perceptual system depends on the active tuning of the tensile state of soft tissues-a medium in the sense of both physics and biology where mechanoreceptors are embedded-so as to separate informative invariants.An electromagnetic micro-sensor was attached on the nail of each index finger of the participant reading braille.(C) Representation of the three angular displacements in the sensor attached on the nail of the index finger [15].

Epilogue
In this paper, I explored the issue of what kind of opportunities the coincidence of the following two notions of medium in reality can make available for intelligence: (1) the substance in which an organism lives, grows, and travels and (2) the substance that propagates and transmits the energy.Drawing on the studies on sensory ecology and reservoir computing, this paper illustrated a principle how pattern and change that are formed in an environmental medium, under certain conditions, could serve as the reservoir of information that makes available the open-ended opportunities for reaching the equivalent state in the continuously changing conditions.The present paper further highlighted the possibility that the presence of isometric tension distributed throughout all levels of interconnected, multiscale networks of the body can make available the opportunities for an active perceiver to spontaneously perturb the tensile states of the body's hierarchical networks in such a way to separate the invariant patterns that specify the source of mechanical disturbances.

Figure 1 .
Figure 1.Water velocity 1 minute after an 86 mm-long fish (Lepomis gibbosus) swam across the area.Arrows show the magnitude and direction of the flow.Bold arrow indicates swimming direction [7].

Figure 2 .
Figure 2. (A) Examples of individual lateral velocity traces as a function of the dominant reading finger's lateral position on a single line of text (shown below as rendered in contracted braille) of a fast reader (top: In this trial, the end of the line was read with the other hand) and a slow reader (bottom: In this trial, the first word of the line was skipped).(B) Set-up of braille reading experiment of children.An electromagnetic micro-sensor was attached on the nail of each index finger of the participant reading braille.(C) Representation of the three angular displacements in the sensor attached on the nail of the index finger[15].