Sarah Jeong
In his paper, “The Symbol Grounding Problem has been solved. So what’s next?” Luc Steels claims that the solution to the Symbol Grounding Problem has, in a way, snuck into the world of cognitive science through recent developments in AI technology. It seems, however, that in order to understand his reasoning one must adapt a rather modified view of the Problem. Steels claims that the basic building block of cognition is the semiotic triad, which relates an object to a symbol and a concept. A symbol may be deemed ‘grounded’ if there is a method that relates a symbol to the object. Since each object has its own semiotic triad, and since objects may be related to each other in their contexts, many links may form between many semiotic triads to create a giant semiotic map. Steels characterizes these semiotic maps as dynamically expanding and reshuffling each time an individual interacts with the outside world.
The reason why the symbol grounding problem has remained a massive headache to cognitive scientists for the past thirty years is because computer scientists programmed robots with a priori information. Robots were created with connections already established between the different semiotic triads; that is, they came with parts of the semiotic map already made. This is the exact opposite of how human cognition develops. A child learns through human interaction and countless series of trial and error all the fine tuned meanings behind each object and the contexts behind such meanings. A child learns that a hot stove is harmful and that he or she must not touch one if pain is to be avoided. A hot stove conjures images and feelings of fear, pain, and even a sternly said “NO!” Robots with a priori information already encoded come prepackaged with all these reactions and furthermore, they fail to develop any other connections. Even more, humans are the product of millions of years of evolution. Humans instinctively know when they are hungry and almost immediately take action to relieve their situation—a baby cries for milk; a teenager reaches for chips. When the batteries of a robot are running low, they lack the motive do anything unless they are encoded to do so.
Steels proposes that the Symbol Grounding Problem is simple and sweet: autonomy. In this light, Steels states that in order for symbols to be grounded within robots the way they are in the human mind, robots must develop and operate in the same way the human brain does: autonomously. As humans, robots must learn through interaction with and feedback from the world in order for semiotic maps to be made. With the advent of artificial systems that can autonomously set up and coordinate semiotic landscapes, the Symbol Grounding Problem has been solved via autonomously established semiotic maps. Steels additionally poses the question, “how come we humans are ever able to communicate and share thoughts at all?”(Author, Year, Page) He elaborates on the significance of interaction and its crucial role of coordinating individuals’ conceptual repertoires. Such must be the same for robots as well. Adaptation to the environment must be done through primarily gauging success in interactions.
Steels’ proposed solution to the Symbol Grounding Problem makes theoretical sense. If robots are to think and communicate like humans, they must develop the skills to do so in the same manner that humans do. The problem with previous approaches to solving the Problem is the ill-conceived notion that the grounding of symbols is a universal and once-and-for-all affair. If this was the case, the robotic cognitive capabilities in the form of categories and lists will only go as far as the programmer is willing to invest time. Steels says himself that “…the methods needed to ground symbols [is] hugely complicated…”(Author, Year, Page) We can do away with complex pre-encoded lists and maps; the key to grounding symbols is through learning and remembering.
Steels takes some time in his paper to elaborate on the nature of c-representations. As opposed to m-representations, which pertain to those created by human cognition, c-representations were created by computer scientists to stand for something of the physical world. This might be a stretch, but one could suggest that that m-representation resembles the 1st person point of view because it is more personal and that the c-representation resembles the 3rd person point of view because it is more objective. In this context, the 3rd person objective representation seems imperative for higher levels of cognition to be possible. Forms of objective representation include images, sculptures, and language sentences. At some point, external representation-making becomes internalized to form the basis of thought; one begins to think in terms of the external. This suggests that external and internal representations are closely intertwined and dynamically co-evolve with respect to one another. Steels coins this manifestation ‘intelligence with representation.’
Ultimately, if a robot is to successfully communicate and collaborate with humans, than it must mirror human nature. Theoretically, the more a robot’s thought process mirrors that of humans, than the more successful it will be. Steels proposes that the solution to the Symbol Grounding Problem lies in this idea. Representation is one component of human cognition, and is comprised of a dynamic relationship between the 1st and 3rd person. Robots must be made in this manner as well.
References:
Steels, L. (2006) The Symbol Grounding Problem has been solved. So what's next? pp. 1-18.
Steels, L. (2004) Intelligence with Representation. Phil Trans Royal Soc Lond A. Volume 361, Number 1811 / October 15, 2003, pp. 2381 – 2395.
Comments (5)
Nicholas Davis said
at 12:35 am on Oct 6, 2008
The correlation between first and third and c/m representations is an interesting point. It is hard to swallow that naturalistic representation should only happen in the first person perspective because it is kind of like the third person representation is necessary for humans to represent meaning at all. To me, it seems the correlation may exist, but it may be the reverse. C-representations could be seen as 1st person because it is sensory input that can be grouped, data points that can be put into a computational array and be ‘meaningful’ depending on the manipulations to it. Whereas third person representation more closely resembles how humans do it. But I see what you mean with the ‘personal’ nature of first person, it seems like that automatic state would be where emotions and things happen; like gut reactions that don’t take any representations. However, in order for objects in the world to take on any kind of emotional connotation, it would have to be represented in the third person sense. There is some interplay between these spaces that deals development. This could be related to the developmental aspect brought up by Emily and Leland to try to investigate what stages deal with which aspect of the first to third person representations and c/m representations.
Matthew McCroskey said
at 1:58 pm on Oct 10, 2008
I think you've made some great points here, Sarah. I agree with Nick that the exact correlation between first- and third-person points of view and c and m representations is a little fuzzy, but definitely worth investigating further.
I'm a little concerned with how we're going to implement the "autonomy" that both you and Steels (and I, too) think is so vital to emulating human cognition. His basic concept seems to be that humans and other animals have the property that they, without "conscious thought", employ complex behaviors to satisfy their basic needs. This is all fine and good except that, really, it's hard to argue that robots have any basic needs. I suppose we could say that power is a basic need, but unlike humans, as a robot's power runs out, it does not become increasingly sluggish. It simply works perfectly until its power source is removed, and then its dead forever. Perhaps we could emulate "low blood sugar" or the basic equivalent in our robot by writing a script that makes the robot operate less efficiently as its battery runs down? While this would emulate human hunger in a sense, I still don't see how it (or any equivalent) is going to give the robot the sense of urgency that humans feel about having their basic needs satisfied. Hopefully, we'll have an opportunity to discuss this further in class today.
Nicholas Davis said
at 2:33 pm on Oct 10, 2008
The low blood sugar idea is an interesting one, but in addition to that, we would have to give it a goal of optimum functioning in order to make it care that it is doing things poorly. Matt's comment to this is: why would you hamper your agent? I said to this that we want to make it more naturalistic, that it needs to complete certain goals in order to continue functioning, therefore creating a sense of autonomy. Another idea is to have a regular 'exam' for the robot, for example if it is in the scenario that makes a socailly constructed meaning, give the robot a test of definitions to gauge what it has learned relative to the other agents, the higher the score, the better. But this would require automatically encoding a desire to want to have its actions perfected. Which wouldn't necessarily be a bad thing because then it might keep trying different action routines in order to optimize efficiency.
Curtis Holmes said
at 3:20 pm on Oct 10, 2008
In the second paragraph, you talk about how in the past robots have been developed with “parts of the semiotic map already made.” I remember Mark Turner in COGS 201 telling us about how he did AI programming back in the day about how they would do exactly this; they would manually put in “chair” and then type associations to it like “legs”, “back”, “sit”, etc. This failed, obviously, because the robot isn’t learning. However, there must be some sort of prepackaged knowledge for the robot, sort of like instincts for humans. Human babies know to cry when they are hungry, but they don’t know that a stove is hot and hurts. So, where do we draw the line? What can we give the robot without ‘cheating’?
maxjensen said
at 12:22 pm on Oct 17, 2008
On Matt and Nick's "low blood sugar":
Per Aage often likes to say that a perfect robot is useless, we need one that can stumble around and make mistakes. Stumbling is very related to playing, in the sense that it teaches the body what motor commands are needed to balance.
Sarah,
This is a solid, simple statement of the problem, which is great. We need the robot to _learn_. I then also relate your ideas about 1st and 3rd person representation back to Emily and Leland's qualm what it is to have a representation, and I agree with them that we need Theory of Mind.
You don't have permission to comment on this page.