"Therefore, I believe that utilizing the 3rd person representation should also be tapped into only after the basic principles of communication are developed" This spawned an interesting idea. What came to me was a gradual increase of usability of the third person perspective. For example, when the agent is created, it will have a third person representation, but maybe it won’t know how to do any work with it. Then gradually, it could learn that by interacting with the environment, this third person space changes, or it can decouple this space and plan things out. All these advances coming from gradually more complex cognitive operations.

Going along this same thread, and related to sarah's paper (the relationship between c/m representations and 1st to third person representations), there is most likely some developmental milestones relating to the use of the first to third person representation. It would be interesting to take a look at Leland's list to see where the third person representation begins, how it is utilized in the different behavioral stages. A nice diagram tying all these ideas together would be optimal. It would help structure the computational aspect in a developmental manner.

I have a potential algorthium for dealing with attention. It makes a few fundamental assumptions, but it does work.

What are the specifics of that algorithm Matt? You started explaining it to me the other day, but maybe it would be good to describe what is actually going on. I'll say what I remember. Basically, this algorithm deals with finding the probability of input to be in some underlying 'state.' For example, the agent could be viewing the actions of another, this would essentially be string of inputs, the state could correlate with what the other agent is 'attending' at that moment, and this algorithm would give some probabilities for what the other agent's attentional state could be. However, would we then have to state every possible thing that the other agent could be attending to or what. I'm still not sure on the specifics of this algorithm.

The algorithm is called mixture model, and it's used to describe the probability of some state causing some action, or connecting the causing and effect. The algorithm tries to create an association between a one dimensional input (ie: a series of numbers) and a number of states. The algorithm attempts to create some kind of association between the two defining the states as the median of binomial distributions, such that when a specific state is active, the output of that state is the numbers which fall inside said binomial distribution. In order for the algorithm to better approximate the relationship between the states and the one dimensional input, it uses gradient descent.
In application of this algorithm, the input could be the how shiny an object is and states can be pay attention to it and do not pay attention to it. It can be applied to shared attention where the input would be various numbers representing the actions of another agent and states can be should the robot pay attention to these actions or not?
One of the fundamental assumptions here is that because the algorithm does not know the correct associations that it should make, it can and does get the association wrong even if the correlation is correct. To ground this in humans would be where someone associates hugs as being repulsive, it's an incorrect association.
I'll be going more indepth of this topic on friday.

I understand and absolutely agree with the basic premise of your argument, Emily; I think that, if we want to meaningfully emulate human cognition, we should consider to the greatest degree feasible the means by which we acquire human cognition. As you mentioned, trying to build some sort of human-like intelligence from biological first principles is certainly out of the reach of this class, if not out of reach altogether. Attempting to mimic the language learning process of human beings seems more manageable, but is still quite daunting. Also, we seem to have been focusing on a robot which has a human-like capability to navigate more so than a human-like capability to communicate. On the one hand, I'm not sure that it is feasible to try to create within this semester a robot that can do both simultaneously. On the other hand, though, it could be argued that, if we can create a general model of human cognition, navigation and communication would be rather naturally intertwined.

I'm a little confused about the algorithm Matt describes, but feel confident that I'll better understand it and its relevance to implementing the concepts of your paper after today's class.

Emily, your reasoning makes a lot of sense to me. I think that before we even go into the details of word lists and semiotic maps, we have to think long and hard about how humans acquire cognition and how we can emulate the same process in robots. Attaining cognition is a long and complicated process and you are right that we should not bypass this critical component of development. In my paper, I mentioned that part of the reason why the symbol grounding problem is a problem to begin with is because computer scientists pre-encoded robots with a lot of essential knowledge. You and I share the view that if robots are to think and communicate like humans, they must develop the skills to do so in the same manner that humans do. Humans learn meanings behind things through a lifetime of trial and error and through gauging the success of interaction with the outside world. If a robot is pre-encoded with information, than there is no way that the robot can understand the meaning and significance of that information. To me, this a prime source of the Symbol Grounding Problem.

Yes, this pre-encoding is precisely the thing. The alternative to this completely pre-encoded thing is starting with some schematic layout, like image schemas, or something to that effect, which is what the work of Paul Cohen was trying to do. He would give the robot certain schemas like 'container' and then have different values which that thing can have 'empty,' 'full,' 'contained' and also there would be slots within that schema to further delineate the feature of this object. Like the closeness schema could be part of the description of container schema, for example a container could have the value of full and the have the closeness as a another image schema within the container schema, which in turn would have a value of close or far. But, this approach still runs into the same basic problem of grounding these original lexical items. It is a nice idea of starting with some schematic knowledge that has interchangable parts, but there is still something missing. This is where perceptual symbols have something to say. The thread on agents sharing third person representations mentions this also. Which perceptual symbols deal with making meanings based on sensory input, recording certain portions of sensory information and relating it to the experience that provoked it. More information on perceptual symbols on the readings list and a summary here: http://cognitiverobotics.pbwiki.com/General+Project+Proposal (in appendix B).

I agree with Emily about the symbol grounding problem. We should try and stimulate the robot as we do with children. I also like the ideas you (Emily) present about how we should first deal with communication and not with the 1st and 3rd person issue. I think you would like the diagram I will present (I think it will help us all understand what needs to be done before we go further).

Emily,

This is a very nice idea. The reason why studying development is so helpful (for all things cognitive and psychological) is that we can see what's missing from the child's mind that makes it a child's mind (and not and adult's). I was thinking about Jean Piaget's stages, which I know are quite controversial, and I'm certain many corrections have been made to them. However, there certainly are stages (however poorly defined they may be). In Piaget's model, children start with only meaning as physical. They then move to linguistic repetitions, and so fourth.

Using models of actual human cognition, including the metaphors used in the field, will allow Robotics to speak with the other domains, which I think is what I like about your paper.

Max