George Kyriakopoulos
SPM - Independent Study – 490
Prof. Tony Chemero
Explorations In Neurosemantics using Computer Models of Neural Networks, Embodied and Evolved in Virtual Organisms
Framsticks 3D Artificial Life is a three-dimensional life simulator program designed to simulate organisms modeled as mechanical structures (bodies) and control system (brains). The objective of the program is to study a simulated evolutionary process in a three-dimensional artificial world with artificial organisms. Genotype representations of the creatures are descriptive of both the physical structures (body) and neural networks (brain) of each individual organism (figure 1-2). The basic structural elements of the simulated organisms are rods or sticks (Framsticks), utilizing a finite element approach (figure 3). Concerning the “brains” of the artificial creatures and the set neural parameters defined by the program, there are three properties influencing the behavior of each individual neurode (artificial neuron), and thus the creatures as a whole:
- Force, value range: 0-1, default 0.04
- Inertia, value range: 0-1, default 0.8
- Sigmoid, value range: any real number, default 2.0
Force and Inertia influence changes of the inner neuron state. In each simulation step, the neuron state is modified towards the value calculated from input excitations. Force determines how fast the value is changed (maximum value of 1.0 gives instant reaction, while low values (default) cause a smooth charging and discharging of the neuron. The neurons Inertia is similar to the physical inertia of a body: sustaining its state change tendency (low inertia values having little influence, while near maximum values (1.0) resulting in oscillation of the neuron state. The Sigmoid coefficient changes the output function (figure 4). The evolutionary process implemented in Framsticks is designed to mimic as closely as possible a natural environment while also allowing for manipulation by the user. Genotypes to be tested are selected from a predefined gene pool (figure 5). The specific selection methods are defined by the user, choosing between random, fitness-proportional (roulette) and tournament (between competing genotypes). Fitness is defined by the user in terms of the desired variables (velocity, distance traveled, consumed energy etc…) and is monitored by the program through organism histories. The simulated group or individual organism is termed the “population”, with all properties continuously measured for each program step.
Due to Framstick’s versatility and user-friendly design, it can be utilized as a research tool for various disciplines including not only those of the sciences but also those of the humanities. A recent article by Professor Pete Mandik of William Patterson University of New Jersey, “Varieties of representations in Evolved and Embodied Neural Networks”, used the program in an attempt to answer or at least shed light on what are understood in the Philosophy of Mind as questions of neurosemantics: How do neural states of organisms have representational content and how did they evolve to be so? The goal of the study was to attempt to understand, in the simplest terms, the conditions required for mental representations to exist in evolved and embodied neural networks. If brains are made of neurons, and minds are made of mental representations, then there must be some way to create mental representational states from neural states. It seems, however, that this notion has led Mandik to making certain assumptions or claims based purely on hypothetical findings.
Mandik initially explains a distinction between two types of mental representations: those of indicative and imperative representations. Indicative representations can be understood as perceptions and memories that are brought into conformity with the world; taken in from our environments as stimuli or experiences. Imperative representations can be understood as intentions of which the world is brought into conformity with; those brought forth and put into the world. Mandik further explains the problem of showing that theories of representational content are compatible with assumptions about the roles of representations within a causal economy: The Economy Problem. Decomposed, we are left with a question of representational content and representational vehicles. Are the conditions establishing representational content for perceptual representations the same (quantitatively and qualitatively) as those establishing such mental states for memories and intentions – are distinct conditions necessary? Are the vehicles for perceptual representations the same (quantitatively and qualitatively) as those for memories and intentions – are distinct vehicles necessary?
Mandik proposes a reverse strategy for answering these questions. Rather than attempting to create theories for complex human brains/minds, one must start with simple organisms, describing how their survival promoting behavior is accomplished, and work backwards – a naturalization of representations. In this sense, if we are to accept the Framsticks program as a viable source for studying such simple organisms, we can see the inherent potential for such a research program to provide important results. Mandik characterizes this approach as the Animat Approach (animat = artificial animal), utilizing three explanatory strategies:
Synthesis - element of explaining target phenomena through attempting to synthesize artificial versions of organisms - inherited from GOFAI and Connectionist approaches.
Holism - focusing on the entirety of an organism far simpler than the human case – understanding in the context of the organism, embodied (the whole) and embedded (in an environment).
Incrementalism - concept of building from the simplest to the more complex via gradual addition of complicating factors.
In utilizing Framsticks for this approach, Mandik attempts to show that simple autonomous agents with few neural controllers and neural connections are capable of acquiring and sustaining, in an evolutionary context, several varieties of mental representations. In this sense, it may be possible to identify what the simplest possible conditions of neural complexity would be, allowing for the most minimal of cognitive behavior.
Mandik defines the four creatures of which he researched as:
Creatures of Pure Will - creatures with no sensory inputs.
Creatures of Pure Vision - creatures that directly perceive some environmental properties.
The Historians - creatures that employ a memory mechanism that allows the comparison between a current stimulus and a remembered stimulus.
The Scanners - creatures that infer or compute the locations of environmental properties based on a comparison of sensory representations of the environment and representations of the states of their own bodies and actions; a form of Action Oriented Representations.
The creatures of Pure Will described by Mandik may be understood as the simplest of embodied and evolved Framsticks artificial life creatures. Their existence is based on the performance of a Central Pattern Generator (CPG) modeled in their NN "brains"(figure 6). Without any sensory input, these creatures move through their artificial environment solely in virtue of the oscillating signal originating from the CPG. The signal is sent through the remaining network to the muscle neurons creating a very simple pattern of locomotion. Mandik claims the outputted signals from the Central Pattern Generator to be the first instances of any representational content; what he calls procedural representations. He further explains that as the CPG gets more complicated in terms of the number of neurons designated for the recurrent network, a more "appropriate motor representation of the ideal configuration of bodily motions that will propel the creature forward" can be achieved. The outputted representations, as defined by Mandik, are explained as "representations with imperative contents; contents with success conditions instead of truth conditions".
The creatures of Pure Vision are slightly more complex in that they have the addition of some sensory input to their NN. Mandik explains these creatures as utilizing sensory input to exhibit taxis: motion towards or away from stimuli (directional - ex. positive photo taxis), and kinesis: motion triggered or suppressed by stimuli (action caused by the presence or absence of specific stimuli). He provides as an example of such a creature, a four-legged "food finder", whose NN "brain" consists of the standard three neuron CPG driving the limb muscles for locomotion and a detached stimulus orientation network consisting of two sensory inputs sending their output to a single steering muscle. The two sections of the NN (CPG: limbs , sensory input: steering muscle) are not connected to each other and perform separate tasks. Through the use of this two sensor stimulus orientation network Mandik explains the creature as having "neural states representative of two dimensions of spatial location relative to an egocentric reference frame": the first sensor representative of how near or far the stimulus may be, and the second sensor representative of to-the-left or to-the-right spatial location. I have provided a diagram of what such a creature’s NN would look like, provided I have made some alterations to the simpler CPG used by Mandik (figure 7). Mandik further investigates and compares the performances of creatures with varying numbers of sensors (from none to a maximum of three). He concludes that creatures with sensory input are capable of representing as many dimensions of spatial location as they have sensors (two sensors - two dimensions), and that those with more sensors generally do better in more diverse environmental conditions (land and water). Mandik explains that even creatures with only one sensor still have some advantage over other creatures with no sensors: “being able to detect a single dimension of proximity can thus allow the creature to stop long enough so as to enjoy the meal...." [the creatures I have observed, regardless of the number of sensors, never stop moving, nor do they ever fully consume the energy balls in one sitting]. In challenging his own hypothesis, Mandik makes a point to distinguish between the creatures of Pure Vision utilizing pure sensory representation only, and those to be introduced shortly. He explains that the single sensor creatures which he explains as utilizing memory and active scanning to perform the equivalent tasks of those with more than one sensory input, should not be held to the hypothesized, number of sensory inputs = number of represented dimensions.
The Historians are defined as creatures implementing memorial representations. If a requirement on representations is to represent the spatially and/or temporally remote, such creatures utilizing memory rather than multiple pure sensory inputs would better satisfy this constraint. Mandik explains the simple recurrent networks used for the CPG’s as being sufficient for providing the added service of a short-term memory store. Two of the three components required for memory: encoding and maintenance, are said to be supplied by the recurrent network envisioned for such creatures, however, the process of implementing retrieval still posses as a challenge. Inspired by testing of memory in bacterial chemo taxis of E. Coli, Mandik proposes an analogous mechanism for the comparison of present and past stimuli. Through implementing a delay in a single sensor stimulus orientation network, Mandik attempts to offer a memorial solution to taxis. In sending a sensory signal through two connection channels, one with multiple neurons in series leading to the steering muscle, and one with a direct connection, Mandik believes the comparison of past and present sensory input to be possible. If this is so, a creature with only one sensory input can use this comparison of past and present to “represent the egocentric location of the stimulus in two dimensions” rather than only one.
The Scanners are the most complex of creatures discussed by Mandik. These creatures represent an attempt at implementing the capacity for action-oriented representations into the NN of the simulations. The bodies of the creatures are designed so as to have a long snout-like limb working as an oscillating scanner. The scanning is what is thought to allow for the building of two-dimensional representations. Unlike all previous creatures, the stimulus orientation network and the CPG are now not separated. Mandik explains that the stimulus orientation network receives input from not only the single sensor but also receives information as “feed-back” from the muscle that controls the scanning motion (figure 8). It is thought that this configuration may encode information through receiving sensor activity as proximity information for one dimension, and comparing sensor state to muscular feedback for the second dimension. In other words, the sensors would pick up the near/far-ness of the stimulus while the sensory/muscle comparison would gain to-the-left or to-the-right information: if sensor activity is high and the snout is bending to the right, then the food is to the right. Similarly, creatures with this NN configuration can gain two-dimensional representation through the comparison of sensory activity to an efference copy of the command signal sent to the scanning muscle. This is an almost identical configuration with the only difference being the substitution of the muscle feedback input into the stimulus orientation network with an efference copy input of the signal (figure 9). The creature now compares the sensor activity with an efference copy of the signal sent to the scanning muscle.
All of the creature codes provided by Mandik for this research are structured as Scanners. These genotype encoding were easily altered so as to provide observable examples of the other three types. Those in the Muscular Feedback and Efference Copy groups all have the long snout-like limb as earlier defined, and have stimulus orientation networks that receive an input from the CPG. The NN layouts of the two groups of creatures: muscular feedback and efference copy, are almost identical, with the only difference being in the efference copy group, in which one neuron which had been left unused in the muscular feedback group, is now connected to a muscle. In the muscular feedback group, this muscle was receiving its input from a neuron sending output to four other neurons (figure 10). This, I am quite sure, has no effect on the creature’s action. Without looking at detailed images of the exact placement of the various neurons on the body of the creatures and the various connections between these neurons, no difference between the groups can be noticed. The placement of the important elements such as the smell sensors and steering muscle are identical throughout all four creatures. Through inspection of the exact connections of the NN's, we can see how the two groups of creatures may differ (figure 11). The majority of the connections are the same, with differing weights, except for a string of four neurons in each group. This string of neurons is located in the creatures' stimulus orientation network and makes up the parallel hidden layer found between the sensory input and the steering muscle. After close inspection, it becomes evident that these differing connections are a result of the different placement of the neurons on the creatures' body, having no effect whatsoever on the creatures behavior. This consideration brings me to questioning the validity of Mandik’s claim of the creatures utilizing different forms of information (feedback and efference) in making their comparisons. It seems to me that these two groups of creatures are working with the same information, in the same way. I see no difference between the efference copy group and the muscular feedback group other than the connection weights. If the creatures are in any way functioning differently, it can only be due to these weights, which is quite doubtful.
The notion of muscular feedback as defined by Mandik and utilized in the creatures’ brains seems to be somewhat problematic. Inspection of the NN of the two Muscular feedback creatures shows that the only instance of a recurrent network feeding information back into the system is that of the initial CPG. The continuous signal sent to the stimulus orientation network from the CPG with the addition of the sensor reading shows me nothing of a feedback circuit. I believe a better classification of these creatures would be muscular input rather than feedback. The stimulus orientation network is receiving an input from the CPG, which it seems is what Mandik explains as being the muscular feedback directing the orientation of the steering muscle.
I have come across a certain difficulty in attempting to decipher what exact part of the creature’s body has the task of directing itself towards food sources. In utilizing the oscillating scanner it seems clear how the creature could pick up information of food sources in its environment and in theory then move towards it, but I fail to see how exactly this snout like stick is meant to propel the whole to the source. I have come to the conclusion that the “snout” exercises no significant moving force onto the body. Through observation of the creatures actively traversing their artificial terrain I was unable to see any clear-cut evidence showing me the moving force of the "snout". The creatures seem to follow a random circular pattern of locomotion while the "snout" swings back and forth, causing no real or apparent change in direction. Furthermore, when in close proximity to a food source, the "snout" shows no apparent attraction to any specific area, continuously swinging back and forth, while the body passes right by the source, seemingly ignoring the chance to eat.
The claims made by Mandik must be purely hypothetic. There is no empirical evidence pointing towards the validity of his claims on the representational capacities of the Framsticks creatures and their NN’s. After detailed observation of the creatures and further study of the program itself, I find no reason to accept that these artificial creatures have the ability to create mental representations of their environment, memorial or otherwise.
Figure 1. Sample genotype encoding
Figure 2. Sample NN brain
Figure 3. Sample bodily structure (parts and joints) including physical forces calculated by Framsticks physical simulator.
Figure 4. Formulae describing input: weighted sum of neuron inputs
the work of the N neuron (X) velocity: analogous to physical velocity
state: internal state (analogous to physical location)
output: output signal
Figure 5. Evolutionary process in Framsticks 3D
Figure 6 - Creature of Pure Will
Figure 7 - Creature of Pure Vision
Figure 8 - Muscular Feedback Creature
Figure 9 - Efference Copy Creature
Figure 10 - Comparison of Muscular Feedback (left) and Efference Copy (right)
Muscular Feedback:
llfffX[0:1.554,2:-2.498,1 :-0.990][-1
:0.676,0:1,0:0.822][-1:0.662](RRllfffMMMX[|-1:-686.752]llFFFMMMX[|-2:-0.832]
,LfffIXllfffMMMX[|3:0.676,4:2.068,5:1.341,6:0](RRllfffMMMIX[|-4:0.859]llFFFM
MMIX[|-5:-1.469],LLLLLLLLLLLfffffMMMMMX[4:-1.344,5:124.814][3:-1.986,4:-1.372][2:-1.404,3:0][1:-557.364,2:4.132][S:932.538][|-11:-3.250],RRllfffMMMIX[|-
12:97.620]llFFFMMMIX[|-13:-0.816]),RRllfffMMMX[|-14:-1.927]llFFFMMMX[|-15:-1
.723][-16:0.892])
Efference Copy :
llfffX[0:0.855,2:1.074,1 :-1.696][-1
:0.676,0:1,0:0.822][-1:0.662](RRllfffMMMX[|-1:-686.752]llFFFMMMX[|-2:-0.832]
,LfffIXllfffMMMX[|3:2.784,4:-2.392,5:0.616,6:0](RRllfffMMMIX[|-4:0.859]llFFF
MMMIX[|-5:-1.469],LLLLLLLLLLLfffffMMMMMX[4:-0.940,10:1.186][3:1.598,9:-0.683][2:2.996,8:2.540][1:0,7:-899.612][S:2.632][|-11:-3.781],RRllfffMMMIX[|-12:9
7.620]llFFFMMMIX[|-13:-0.816]),RRllfffMMMX[|-14:-1.890]llFFFMMMX[|-15:-2.282
][-16:0.719])
Figure 11 - Comparison of Creature Codes for Muscular Feedback and Efference Copy Creatures
