The author delves into the microdomain of a tabletop. His Tabletop program attempts to build relationships between objects and groups of objects. For example, Henry is sitting on one side of the table and points to his glass. Liza is sitting on the other side of the table and must point to the closest thing that resembles a glass. If she only has a cup on her side of the table, does she point to her cup, or does she reach over and touch Henry’s glass? By examining a microdomain in depth, the author hopes to gain insight into many of the problems encountered in everyday analogy-making.

Chapter 1 discusses at length the concept of slippage. For example, a glass is not a cup, but they both slip to the same concept, a container for holding liquid. However, slippage can become complicated when we talk about relationships between structures. For example, if Henry has one plate and two forks and Liza has one bowl and one spoon, and Henry points to the rightmost fork, to what does Liza point? The author makes the point that analogies are often made between objects with regard to their surroundings.

Chapter 2 describes the Tabletop domain and gives the rationale behind the examination of a microdomain.

Chapter 3 describes the Tabletop architecture, which consists of a Slipnet that models long-term memory and a Workspace that models short-term memory. Codelets are used to examine domains and to build structures and relationships. Codelets activate other codelets. The notion of temperature is used to determine when to stop examining the scene.

Chapter 4 presents an example of how Tabletop works and points out several features. Chapter 5 presents the results from two thousand runs. This is the linchpin of the book, according to the author. Chapter 6 compares Tabletop to other historical efforts in artificial analogy-making. Chapter 7 offers conclusions.

The author has laid out his work nicely. He makes the point that prior works have examined sameness between structures, but that these structures have been hard-coded into the applications. Tabletop, on the other hand, is designed to build structures dynamically and then examine them using stochastic processes. Tabletop is a blend of connectionism (bottom-up) and semantic nets (top down).

I found it interesting that, while Tabletop emphasizes relationships between structures, experiments with people reveal that they often make analogies on an object-to-object basis. The author’s exhaustive examination of various situations effectively demonstrates where Tabletop will cross over to an object-to-object comparison. Another interesting point about Tabletop is that it can examine the same tabletop setting twice and come up with two different analogies. Tabletop uses independent parallel processes, which can produce different results.

I found the references to be adequate, but biased toward the author’s mentor, Douglas Hofstadter. I would recommend this book to AI researchers who are interested in modeling human cognition.