The Emotion Machine

Trans-Frames: To represent the effects of an action, it is convenient to use pair of semantic networks to represent what was changed. This is what we did in §5-8 to imagine replacing the top of an arch. This way, one only needs to change the name of a single relationship—instead of altering thousands of points to change a visual picture-like image.

A Trans-Frame for changing the top of an arch

We use the term “Trans-Frame” to name such a pair that represents the conditions before and after some action was done. Then we also can represent the effect of a sequence of actions by linking together a chain of the Trans-Frames to form a story or narrative. Here is a sequence of trans-frames for giving a book:

Such a sequences can describe a script that includes any further details that one might need.

A Script for transferring a Book

Each of these types of representation can answer certain types of questions—but what could enable computers to produce answers so quickly as human brains do? When someone says ‘apple,’ you seem to almost instantly know that a typical apple grows on a tree, is round and red, is about the size of a human hand, and has a certain texture, flavor and taste—yet almost no time seems to elapse between hearing that word and then becoming aware of such things.

To explain how that information could so quickly appear, I conjecture that much of such knowledge is wrapped into structures called Frames. The simplest type of Frame consists of nothing more than a labeled list of some properties of particular object, and you can think of this kind of list as like a printed form that has ‘blanks’ or ‘slots’—each of which can be filled-in with a link to some fragment of knowledge. Then, when you know which slot to inspect, you can quickly retrieve that fragment of knowledge, without need much time to search for it.

A Frame for an Apple’s Properties

Default Assumptions: A valuable feature of a typical frame is that every slot comes already filled in with some ‘default’ or typical value for it. Then you can use such a value to make a good guess whenever you don’t have a more definite answer. For example, you might assume ‘by default’ that an apple is red—but if your particular apple is green, then you will replace ‘red’ by ‘green’ in its color slot. In other words, a typical frame describes a stereotype whose ‘default assumptions’ are usually right—but which we can easily change or replace when we encounter exceptions to them.[173] This would seem to be an important aspect of how we do commonsense reasoning.

Picture-Frames: Every slot of a property list is directly connected to the name of that frame, as in the list above for describing an Apple. However, other more complex kinds of frames may have more connections between their various slots. For example, to represent some view of a room, we could use what we call a “picture frame” to represent each wall of that room by in terms of a few large spatial regions as shown below. Then each such region can have some links that describe the objects close to that part of the wall, as well as some links to other nearby regions. This kind of structure would allow us to do a good deal of commonsense spatial reasoning. [See §§24 of SoM.]

Frames for Including Additional Slots: It makes sense to allow each Frame to include some additional slots for representing knowledge that is not already described by the networks contained inside that frame. For example, here is a Trans-frame for Joan’s trip to New York:

This frame includes two semantic networks that describe the situations Before and After that trip was taken. However, it also contains other slots that describe when, how and why Joan took that trip. Then the default assumptions included in those slots can supply additional knowledge for answering such questions as these.

Where did that action occur and when? Who or what caused it to happen?

Was it intentional or not? What purposes was it intended to serve?

What devices or tools were used? What were its other side effects?

Which resources did it engage?    What was expected to happen next?

This suggests an explanation of how we quickly use our commonsense knowledge—without any sense that we’re doing it: it is an example of the “Immanence Illusion” that we described in §4-3.1. As soon as you activate such a frame, then many questions that you might otherwise ask will already be answered before you can ask them—because they are among the default values of that frame’s slots. For example, if you heard that Charles was holding a book, you would not stop to ask why he was holding it; you would simply assume that he has the most usual goal for which any person holds anything—namely, to keep it from falling to the floor.[174]

Connectionist and Statistical Representations.

Student: This book suggests some ideas about how high-level knowledge-based systems could come to achieve things like human commonsense reasoning. But why were no such systems built in the past?

Work on such systems almost came to a stop in the 1980’s because most researchers recognized that that this would need ways to acquire and to organize millions of fragments of commonsense knowledge. That prospect seemed so daunting that most researchers decided to search for simpler alternatives. This led to many attempts to design some single process that would somehow evolve whatever it needed—along with learning all the knowledge it would need by interacting with the external world. Some of these “baby machines” did learn to do some useful things (such as to recognize various kinds of patterns) but as we noted in Chapter §6, none of them went on to develop more higher-level reflective ways to think.

Why were none of those “Baby Machines” able to keep extending their abilities? It appears to me that this failure came mainly because most of their designers decided that their systems should represent the knowledge they were to acquire mainly in numerical terms. Consequently, most of those ‘baby machines’ were designed to use the techniques called Neural Networks, Perceptrons, Fuzzy Logic systems, and “Statistical Learning Programs.” All such systems represent knowledge about the relations between things by assigning numerical ‘weights’ or ‘strengths’ to connections inside a network of nodes. Such a representation might look like this:

Here we see only one kind of link, which reduces every type of relationship to a single numerical value or