"Evolution of Communication and Language Using Signals, Symbols and Worlds" is by A. Cangelosi and it has a 2001 publication date.
It is available at CogPrints, IEEE and CiteSeer.
This is an important paper because I am hoping that it will answer my questions about the paper, "The Adaptive Advantage of Symbolic Theft over Sensorimotor Toil: Grounding Language in Perceptual Categories" by A. Cangelosi published in 2002.
1.0 Introduction
1.1 Icons, Indices and Symbols
Studies on communication systems distinguish between icons, indices and symbols. Icons have physical resemblances to the objects they represent. Indices provide associations between time and space to objects. Symbols emerge from social conventions.
There is a distinct gap between animal and human communication systems.
In animal communication, signals relate to world entities and associations arise from instinct and conditioning. In human communication, signals relate to world entities and other symbols.
1.2 Signals, Symbols, and Words in Language Evolution
Different stages of semiotic complexity include simple associations between symbols and objects, and associations between symbols.
I do not fully understand what that means but I get the thrust of it.
Sensorimotor grounding helps organisms.
I like the diagram that describes the differences between indexed, grounded and non-grounded languages. The work by Savage-Rumbaugh on ape languages is an inspiration here. Savage-Rumbaugh's work looks interesting.
The papers I read on evolutionary emergent communication were only on simple forms of communication.
2.0 Evolution of Communication Using Signals
2.1 Evolution of Signals - Model Setup
The mushroom world model comes from an older paper on Econet models.
I hope this paper answers my questions about the mushroom world model. One of my questions is how does the calling and listening work? Another question is where does knowledge about correct actions that trains the network come from?
I do not like how the fitness formula abstracts the process of surviving.
I still do not understand whether adaptation occurs because the neural network changes because of an individual's learning or because of genetic mutations.
Okay, after a bit of thought, I understand more about how and what type of learning occurs. In the initial population, the experiment randomly initialises all foragers. Each forager undergoes individual learning where their neural network adapts via back propagation.
Note that the experiment uses the neural network before back propagation as the genetic code. At the end of a population's lifetime, the experiment picks the top fittest foragers and uses the neural networks of these foragers as it was before back propagation to create the subsequent generation with a small mutation factor.
This is important because the foragers do not pass on their individual learning. An example that I think I repeat later on is how I will not pass on my training from my computer science degree to my child. I will pass on my genetics only.
Okay, here is how the calling and listening system works. Foragers can only perceive mushroom characteristics when foragers are adjacent to a mushroom. Each time a forager moves the system selects randomly another forager to speak to the first forager. The second forager can see the mushroom that is the closest to the first forager.
It feels like there is significant number of things missing from this experiment. Some of the things are interaction, clarification, questioning, answering and reaching a common context. However, this experiment is important because it provides a good starting point where the author has already done significant amount of thinking for me.
The experiment has three populations. The first population has no language. The second population receives a language that does not evolve. The third population evolves its own language.
2.2 Evolution of Signals - Results
The results of the experiment are as follows. Populations with language performed better. Populations that received language and populations that developed language performed similarly.
The foragers have a signal based communication system rather than a symbol based communication system. Remember that signal based communication is where signals refer to a world object and that symbol based communication is where symbols refer to world objects and other symbols. This does not seem like the best definition but it is sufficient for now.
Things like time are important because they aid the listener to determine when a speaker has finished a sentence. I think constraints such as time and the sounds a speaker can make and sounds a listener can hear can potentially lead to the same symbols for multiple things and multi-symbol utterances. I do prefer agents to be predisposed towards making human like sounds and I think that this is possible to do. Another idea is to have a memory system that records everything and which agents can revisit to enable conceptual revolutions.
3.0 Evolution of Symbolic Communication
3.1 Evolution of Symbols - Model Setup
The population consists of 80 foragers.
There are six types of mushrooms. There are edible mushrooms that are small, medium and big. There are inedible mushrooms that are small, medium and big.
The world is 100 by 100 cells.
At the beginning of each epoch, there are 1200 randomly distributed mushrooms.
There are 200 mushrooms from each category.
A binary string represents the 18 features of each mushroom. A set of three features set to one, identifies the mushroom type.
Foragers receive a point for correctly categorising mushrooms and lose a point for incorrectly categorising mushrooms.
At the end of a population lifetime, the system selects the 20 fittest organisms to reproduce four offspring each. The genotype is the neural network of a forager. The system applies a 10% mutation.
The neural network architecture is as follows. The network is a feed forward type and it has three layers. Three units in the input layer encode the location of the closest mushroom. Eighteen units in the input layer encode the features of the closest mushroom. Eight input units encode communication symbols. The hidden layer has five units. Three output units control movement and identification. Eight output units encode mushroom names.
Symbolic output units fall into two clusters of competitive winner takes all units. Only one unit in each cluster can activate. One cluster has two units and the other cluster has six units. The effect of this is that foragers use two symbols at a time to refer to mushrooms.
What is the significance of this? Something does not seem right here because there are automatically two clusters.
I suspect that foragers will adapt to use the six-unit cluster to refer to mushrooms and to use the two-unit cluster to refer to what to do.
Generations 0 to 299 evolve to be able to discriminate between the six types of mushrooms. These foragers do not communicate.
Generations 300 to 399 do communicate. The 80 new foragers live together with 20 parent foragers. Only the 80 child foragers forage and reproduce.
During each action, parent foragers receive an 18-bit string and produce two output symbols to describe the mushroom. I assume that the calling and listening process is similar to the previous model.
Parent foragers pass the two output symbols to child foragers. Child foragers also receive an 18-bit string 10% of the time. The effect of this is that child foragers need to rely upon parent forager output because child foragers rarely get to directly experience mushrooms.
I wonder if grounding still occurs here because child foragers have not really developed any initial categories by themselves.
I get the listening but not the naming and imitation tasks.
After a bit of thought, some ideas arose about what the listening, naming and imitation tasks actually are. The listening task is what action the child network makes after receiving symbols from the parent network. The naming task is what symbols the child network makes after receiving symbols from the parent network.
The imitation task is how the child network adapts to produce symbols similar to those from the parents. The child network receives symbols from the parent network as input. The child network compares its symbol output with symbols from the parent network. The child network adapts itself so that it produces symbols similar to the parent network when it receives those symbols from the parent network.
Foragers use an error backpropagation algorithm to learn and they use the two symbols from parents as teaching input.
I do not quite get this.
I would prefer to have real error in communication rather than abstractly adding a plus-minus 0.5 error value.
I do not quite get how the parent and child foragers interact but I like the gist of it.
I am also beginning to understand what agents do and do not pass on to the next generation. Agents pass on neural network weights prior to back propagation. Agents do not pass on neural network weights subsequent to back propagation. This is an important difference. It is similar to how you might pass your genetics to your child but you do not pass your knowledge about the world to your child directly. It is all becoming much clearer.
Was there any learning in the previous mushroom world model where the second population received a language?
How is the model for evolution of symbols different to the model for evolution of signals?
A possible avenue for research is to place constraints on the number of calculations an organism is able to do. For example if there are 80 types of mushrooms and 80 signals, it might be hard for an organism to compare one by one. I know that this is no problem for a computer however; an organism might develop useful categories and rules of thumbs.
3.2 Evolution of Symbols - Model Results
In the second experiment, Cangelosi was able to coax foragers to develop languages that are more sophisticated. The languages had multiple symbols, which referred to actions and objects.
Savage-Rumbaugh has demonstrated that apes are able to perform similarly.
Cangelosi repeated the experiment for generation 0 to 299 ten times. In nine out of ten times, population 299 reached an optimal level. An optimal level occurs when foragers are able to distinguish between edible and inedible mushrooms as well as whether the edible mushrooms are small, medium or larger.
Cangelosi took the nine successful populations to simulate generations 300 to 399. Cangelosi permitted generations 300 to 399 to communicate.
Cangelosi gave the nine populations a random lexicon and then repeated this with another random lexicon.
Cangelosi grouped resultant languages into two main categories, good and imperfect. A good language uses at least four signals to distinguish between edible and inedible mushrooms as well as edible mushroom sizes. An imperfect language may fail to distinguish between edible and inedible mushrooms, between edible mushroom sizes or both.
Cangelosi further divided these categories into single-signal, signal-combination and verb-noun.
Single-signal languages use only one cluster to differentiate between types of mushrooms.
Signal-combination languages use both clusters to differentiate between types of mushrooms. Both clusters represent mushrooms and actions rather than one cluster for mushrooms and the other cluster for actions.
Verb-noun languages use both clusters to differentiate between types of mushrooms. One cluster represents a mushroom and the other cluster represents whether to approach or avoid.
The results were as follows. In good languages, 9% were single-signal, 27% were signal-combination and 64% were verb-noun. In imperfect languages, 14% were single-signal, 29% were signal-combination and 57% were verb-noun.
The goal of this experiment is to demonstrate that symbol based communication is occurring rather than signal based communication. Signal based communication uses one signal to refer to an object type. Symbol based communication uses one symbol to refer to an object type and to other symbols.
Cangelosi, 1999, describes an experiment to test whether agents developed a signal or symbol based communication. The experiment worked by testing whether agents could use the verb-noun rule. Agents adapted to associate the verb "approach" with two types of edible mushrooms and "avoid" with two types of inedible mushrooms.
The experiment then introduced a third type of edible mushroom and a third type of inedible mushroom. Agents had to adapt to associating the correct action with the correct mushroom type. Agents could no receive direct feedback on verb association.
Seven out of ten populations successfully adapted indicating that symbol based communication was occurring rather than signal based communication. Admittedly, further testing and analysis is required.
4.0 Computational Approaches to the Evolution of Words and Syntax
I did not get a lot from this section. This section does discuss the importance of grounding symbols. It also discusses how this potentially leads to further language development.
5.0 Conclusion
I did not get a lot from this section.
Binh's Final Thoughts
I still want to know how the neural network training works. Everything else is reasonably clear.
I understand how signal based communication works but not how symbol based communication works.
It would be a good idea to prepare a summary of different types of communication as presented in this paper.
I have a hunch that as the capability of foragers increases, so does the language. For example, if foragers could bring mushrooms back, store them, ration them and so on, the required language would also need to grow.
There is no real coming to common understanding of desired context.
The fact that a bestowed language performs as well as an evolved language worries me but at the same time, it is nice to know that an evolved language performs as well as a bestowed language.
No comments:
Post a Comment