The problem of relevance

In chapter 4, Hoftstadter discusses artificial intelligence and the problem of representation. He defines representation as the end product of the process of perception, which is when a set of raw data has been organized into a coherent and structured whole. There are both long-term and short-term representations that hold importance at a given moment during a mental or computational process.

Of course, there are obvious questions as to how this could be implemented in a program so that representations affect a program's decisions in a human-like fashion. For example, a long-term representation could be a heuristic for solving a certain problem, like a crypto problem with 0 as the goal and one of the numbers. However, implementing the representations is not the only problem to deal with.

Constructing representations poses problems because of the problem of relevance and the problem of organization. The problems basically entail deciding what parts of the virtually unlimited number of environmental aspects are to be included in the representation, and how the representation is actually itself represented in the database.

These ideas are interesting to read about because I had wondered how one would develop a program that learns with experience. A program like that would need to have the ability to decide how much to include from its environment, and it would have to know how to represent that data, if it were to create its own representations over time. This is, after all, what people do.

The Eliza Effect

In Preface 4, Hofstadter discusses the Eliza effect and why it had worried him and his team back in the 1980's. The Eliza effect is basically defined, as Hoftstadter says, as: "the susceptibility of people to read far more understanding than is warranted into strings of symbols--especially words-- strung together by computers." In other words, it also refers to how people mistakingly believe that computers understand things, or can empathize with things, like humans far more than they actually do.

It is because of this that Hoftstadter and his team were losing morale. Other AI research teams were being credited with achieving goals in AI development which Hoftstadter believes was unwarranted, at least as much as they got. I can understand his thinking, and although this preface does seem to be a bit on the complaining side, his complaints seem valid.

From reading about his work with CopyCat and Jumbo, it is very obvious that he is aiming to truly model human cognitive processes with his work in artificial intelligence. This is in contrast to many other AI research groups who develop AI with the goal of achieving the image of intelligence with their projects. For some people, it simply is not important whether the computer truly understands the domains it is meant to be dealing with. For example, AI in video games is meant to mimic human behavior, but with as little overhead as possible. For these purposes, truly modeling human cognitive processes is not entirely important for most people, I guess. However, for people claiming to have a computer program that truly understands something like creating analogies, it has no purpose to claim the goal has been achieved if it has not truly been achieved in a way that mirrors a human approach.

Numbo

In chapter three of the reading, Daniel Defays discusses the Numbo program. He explains the goals of the game of Numble, which is to construct a given number, designated as the goal, from a group of other given numbers, using the three mathematical operations of addition, subtraction, and multiplication. This is a lot like crypto problems, except for that division is not used, and that all numbers do not need to be used to reach the designated goal number in the Numble game.

Still, the same basic principles seem to apply to the two different number games, and much of what Defays discusses reminds me yet again of the way in which I attempted to solve my first set of crypto problems. In Numbo, there are Pnodes which basically serve as markers stored in the Pnet meant to be close enough to specific ranges to be of use. In the first example that he walks the reader through, the goal is 114. In this case, the Pnode “100”, which has been programmed in as part of Numbo, is activated due to its relative closeness to 114.

The system seems to model the way that people think when approaching these problems, but even as Defays himself admits, the Pnet (the area storing the permanent network of knowledge that the program holds) does not know and recognize all of the same types of things that a human would recognize.

Temperature

In this section of the reading, Hofstadter continues to discuss the workings of his Jumbo project. The most interesting notion he brings up is how he incorporates a "temperature" into the system. Using the cell analogy, he discusses how the "cytoplasm" may incorporate a temperature, which serves as a quick global feedback to the state of the system. More precisely, it serves as a kind of barometer of the “happiness” of the gloms within this cytoplasm mixture.

The temperature is low, in fact, freezing, if there is only one single happy glom. The freezing analogy is great because the system does not break up the glom due to the fact that the cytoplasm is frozen. As expected, as the temperature increases, the chance of gloms dissolving increases. I just thought it was neat.

Hofstadter also discusses some classic illusions, relating to the fact that humans only perceive one possibility at a time, and citing things such as the vase-or-faces picture. Completely coincidentally, I was also recently shown a video on YouTube of a dancer’s silhouette that is meant to appear to be constantly spinning. However, the illusion is that it can be seen as the dancer spinning left, or the dancer spinning right. This particular example is very strong, because even after perceiving both possibilities, I was unable to make my perception change at will.

http://www.youtube.com/watch?v=uBTvKboX84E