AAAI FSS-13 and Symbol Grounding

At the AAAI 2013 Fall Symposia (FSS-13) 1 2, I realized that I was not prepared to explain certain topics quickly to those who are specialists in various AI domains and/or don’t delve into philosophy of mind issues. Namely I am thinking of enactivism and embodied cognition.

my poster

my poster

But something even easier (or so I thought) that threw up communication boundaries was The Symbol Grounding Problem. Even those in AI who have a vague knowledge of the issue will often reject it as a real problem. Or maybe Jeff Clune was just testing me. Either way, how can one give an elevator pitch about symbol grounding?

So after thinking about it this weekend, I think the simplest explanation is this:

Symbol grounding is about making meaning intrinsic to an agent as opposed to parasitic meaning provided by an external human researcher or user.

And really, maybe it should not be called a “problem” anymore. It’s only a problem if somebody claims that systems have human-like knowledge but in fact they do not have any intrinsic meaning. Most applications, such as NLP programs and semantic graphs / networks, do not have intrinsic meaning. (I’m willing to grant them a small amount intrinsic meaning if that meaning depends on the network structure itself.)

Meanwhile, there is in fact grounded knowledge of some sort in research labs. For instance, AI systems in which perceptual invariants are registered as objects are making grounded symbols (e.g. the work presented by Bonny Banerjee). That type of object may not meet some definitions of “symbol,” but it is at least a sub-symbol which could be used to form full mental symbols.

From Randall C. O’Reilly, Thomas E. Hazy, and Seth A. Herd, "The Leabra Cognitive Architecture: How to Play 20 Principles with Nature and Win!"

From Randall C. O’Reilly, Thomas E. Hazy, and Seth A. Herd, “The Leabra Cognitive Architecture:
How to Play 20 Principles with Nature and Win!”

Randall O’Reilly from University of Colorado gave a keynote speech about some of his computational cognitive neuroscience in which there are explicit mappings from one level to the next. Even if his architectures are wrong as far as biological modeling, if the lowest layer is in fact the simulation he showed us, then it is symbolically grounded as far as I can tell. The thing that is a “problem” in general in AI is to link the bottom / middle to the top (e.g. natural language).

I think that the quick symbol grounding definition above (in italics) is enough to at least establish a thin bridge between various AI disciplines and skeptics of symbol grounding. Unfortunately, I also learned this weekend that hardly anybody agrees on what a “symbol” is.

Symbols?

Photo taken from the Westin hotel. I just noticed that Gary Marcus snuck into my photo.

Photo taken from the Westin hotel. I just noticed that Gary Marcus snuck into my photo.

Gary Marcus by some coincidence ended our symposium with a keynote that successfully convinced many people there that symbolic AI never died and is in fact present in many AI systems even if they don’t realize it, and is necessary in combination with other methods (for instance connectionist ML) at the very least for achieving human-like inference. Marcus’s presentation was related to some concepts in his book The Algebraic Mind (which I admit I have not read yet). There’s more to it like variable binding that I’m not going to get into here.

As far as I can tell, my concept of mental symbols is very similar to Marcus’s. I thought I was in the traditional camp in that regard. And yet his talk spawned debate on the very definition of “symbol”. Also, I’m starting to wonder if I should be careful about “subsymbolic” vs. “symbolic” structures. Two days earlier, when I had asked a presenter about the symbols in his research, he flat out denied that his object representations based on invariants were “symbols.”

So…what’s the elevator pitch for a definition of mental symbols?

The Code Experience vs. the Math Experience

In the book The Mathematical Experience, the chapter on symbols mentions computer programming [1]. But it really doesn’t do justice to programming (aka coding). In fact it’s actually one of the lamer parts of an otherwise thought-provoking book. It’s not that it’s dated—a concern since the book was published in 1981—but that the authors only provide the paltry sentence, “Computer science embraces several varieties of mathematical disciplines but has its own symbols,” followed by some random examples of BASIC keywords and some operators.

cover of the edition I have

As mentioned in “The Disillusionment of Math,” I’ve always thought of programming as different than mathematics. And I almost always choose the experience of thinking in code over the experience of thinking in equations.

But I suspect others think of these as similar activities occupied by mathematical people. Likewise, if a programmer tells somebody that they are a software engineer, the keyword “engineer” can create a response of “oh you must do a lot of math.”

Continue reading

Mechanisms of Meaning

How do organisms generate meaning during their development? What designs of information structures and processes best explain how animals understand concepts?

In “A Deepness in the Mind: The Symbol Grounding Problem“, I showed a three layer diagram with semantic connections on the top. I’d like to spend some time discussing the bottom of the top layer.

Basic Meaning Mechanisms

At the moment, there seems to be a few worthwhile avenues of investigation:

  • Emotions
  • Affordances
  • Metaphors And Blends

Each of these points of view involve certain theories and architecture concepts. Soon I will have more blog posts describing these concepts and how we might implement and synthesize them. Marvin Minsky’s books Society of Mind and The Emotion Machine have many mental agency and structure ideas that may be useful in this context.

A Deepness in the Mind: The Symbol Grounding Problem

The Symbol Grounding Problem reared its ugly head in my previous post. Some commenters suggested certain systems as being symbol-grounding-problem-free because those systems learn concepts that were not chosen beforehand by the programmers.

However, the fact that a software program learns concepts doesn’t mean it is grounded. It might be, but it might not be.

Here’s a thought experiment example of why that doesn’t float my boat: Let’s say we have a computer program with some kind of semantic network or database (or whatever) which was generated by learning during runtime. Now lets say we have the exact same system, except a human hard-coded the semantic network. Did it really matter that one of them auto-generated the network versus the other as far as grounding goes? In other words, runtime generation doesn’t guarantee grounding.

Experience and Biology

Now let’s say symbol grounded systems require learning from experience. A first-order logic representation of that would be:

SymbolGrounding -> LearningFromExperience

Note that is not a biconditional relationship. But is that true? Why might learning from experience matter?

herring gull and chicks

Well our example systems are biological, and certainly animals learn while they are alive. But that is merely the ontogeny. What about the stuff that animals already know when they are born? And how do they know how to learn the right things? That’s why the evolutionary knowledge via phylogeny is also important. It’s not stored in the same way though. It unfolds in complex ways as the tiny zygote becomes a multicellular animal.

Continue reading

What are Symbols in AI?

A main underlying philosophy of artificial intelligence and cognitive science is that cognition is computation.  This leads to the notion of symbols within the mind.

There are many paths to explore how the mind works.  One might start from the bottom, as is the case with neuroscience or connectionist AI.  So you can avoid symbols at first.  But once you start poking around the middle and top, symbols abound.

Besides the metaphor of top-down vs. bottom-up, there is also the crude summary of Logical vs. Probabilistic.  Some people have made theories that they think could work at all levels, starting with the connectionist basement and moving all the way up to the tower of human language, for instance Optimality Theory.   I will quote one of the Optimality Theory creators, not because I like the theory (I don’t, at least not yet), but because it’s a good summary of the general problem [1]:

Precise theories of higher cognitive domains like language and reasoning rely crucially on complex symbolic rule systems like those of grammar and logic. According to traditional cognitive science and artificial intelligence, such symbolic systems are the very essence of higher intelligence. Yet intelligence resides in the brain, where computation appears to be numerical, not symbolic; parallel, not serial; quite distributed, not as highly localized as in symbolic systems. Furthermore, when observed carefully, much of human behavior is remarkably sensitive to the detailed statistical properties of experience; hard-edged rule systems seem ill-equipped to handle these subtleties.

Now, when it comes to theorizing, I’m not interested in getting stuck in the wild goose chase for the One True Primitive or Formula.  I’m interested in cognitive architectures that may include any number of different methodologies.  And those different approaches don’t necessarily result in different components or layers.  It’s quite possible that within an architecture like the human mind, one type of structure can emerge from a totally different structure.  But depending on your point of view—or level of detail—you might see one or the other.

At the moment I’m not convinced of any particular definition of mental symbol.  I think that a symbol could in fact be an arbitrary structure, for example an object in a semantic network which has certain attributes.  The sort of symbols one uses in everyday living come in to play when one structure is used to represent another structure.  Or, perhaps instead of limiting ourselves to “represent” I should just say “provides an interface.”  One would expect that a good interface to produce a symbol would be a simplifying interface.  As an analogy, you use symbols on computer systems all the time.  One touch of a button on a cell phone activates thousands of lines of code, which may in turn activate other programs and so on.  You don’t need to understand how any of the code works, or how any of the hardware running the code works.  The symbols provide a simple way to access something complex.

A system of simple symbols that can be easily combined into new forms also enables wonderful things like language.  And the ablity to set up signs for representation (semiosis) is perhaps a partial window into how the mind works.

One of my many influences is Society of Mind by Marvin Minsky [2], which is full of theories of these structures that might exist in the information flows of the mind.  However, Society of Mind attempts to describe most structures as agents.  An agent is isn’t merely a structure being passed around, but is also actively processing information itself.

Symbols are also important when one is considering if there is a language of thought, and what that might be.  As Minsky wrote:

Language builds things in our minds.  Yet words themselves can’t be the substance of our thoughts.  They have no meanings by themselves; they’re only special sorts of marks or sounds…we must discard the usual view that words denote, or represent, or designate; instead, their function is control: each word makes various agents change what various other agents do.

Or, as Douglas Hofstadter puts it [3]:

Formal tokens such as ‘I’ or “hamburger” are in themselves empty. They do not denote.  Nor can they be made to denote in the full, rich, intuitive sense of the term by having them obey some rules.

Throughout the history of AI, I suspect, people have made intelligent programs and chosen some atomic object type to use for symbols, sometimes even something intrinsic to the programming language they were using.  But simple symbol manipulation doesn’t result in in human-like understanding.  Hofstadter, at least in the 1970s and 80s, said that symbols have to be “active” in order to be useful for real understanding.  “Active symbols” are actually agencies which have the emergent property of symbols.  They are decomposable, and their constituent agents are quite stupid compared to the type of cognitive information the symbols are taking part in.  Hofstadter compares these symbols to teams of ants that pass information between teams which no single ant is aware of.  And then there can be hyperteams and hyperhyperteams.

References
[1] P. Smolensky http://web.jhu.edu/cogsci/people/faculty/Smolensky/
[2] M. Minsky, Society of Mind, Simon & Schuster, 1986.
[3] D. Hofstadter, Metamagical Themas, Basic Books, 1985.