Artificial intelligence (AI), a branch of computer science, seeks to explore human cognitive processes by creating computer models that accomplish the same tasks that humans do (Boden, 2004; Chrisley, 2004). Researchers in artificial intelligence have tried to explain how you recognize a face, create a mental image, write a poem, as well as hundreds of additional cognitive accomplishments (Boden, 2004; Farah, 2004;
Current Issues in Cognitive Psychology 17 Thagard, 2005). In this textbook, you’ll read about research on artificial intelligence
in Chapter 7 (mental maps), Chapter 9 (language comprehension), and Chapter 11 (problem solving). Let’s consider several important topics within the domain of artifi- cial intelligence: (1) the computer metaphor, (2) pure AI, (3) computer simulation, and (4) the parallel distributed processing approach.
The Computer Metaphor. Throughout the history of cognitive psychology, the computer has been a popular metaphor for the human mind. According to the computer metaphor, our cognitive processes work like a computer, that is, a com- plex, multipurpose machine that processes information quickly and accurately.
Of course, researchers acknowledge obvious differences in physical structure between the computer and the human brain that manages our cognitive processes. However, both may operate according to similar general principles. For example, both computers and humans can compare symbols and can make choices according to the results of the comparison. Furthermore, computers have a processing mecha- nism with a limited capacity. Humans also have a limited attention capacity. As we’ll discuss in Chapter 3, we cannot pay attention to everything at once.
Computer models need to describe both the relevant structures and the processes that operate on these structures. Thagard (2005) suggests that we can compare a computer model with a recipe in cooking. A recipe has two parts: (1) the ingredients, which are somewhat like the structures; and (2) the cooking instructions for working with those ingredients, which are somewhat like the processes.
Researchers who favor the computer approach try to design the appropriate “software.” With the right computer program and sufficient mathematical detail, researchers hope to mimic the flexibility and the efficiency of human cognitive processes (Boden, 2004).
AI researchers favor the analogy between the human mind and the computer because computer programs must be detailed, precise, unambiguous, and logical (Boden, 2004). Researchers can represent the functions of a computer with a flow- chart that shows the sequence of stages in processing information. (Figure 1.1 on page 10 shows a simplified flowchart.) Suppose that the computer and the human show equivalent performance on a particular task. Then the researchers can speculate that the computer program represents an appropriate theory for describing the human’s cognitive processes (Carpenter & Just, 1999).
Every metaphor has its limitations, and the computer cannot precisely duplicate human cognitive processes. For example, no artificial-intelligence system can speak and understand language as well as you do, because your background knowledge is so much more extensive (Boden, 2004). Furthermore, humans have more complex and fluid goals. If you play a game of chess, you may be concerned about how long the game lasts, whether you are planning to meet a friend for dinner, and how you will interact socially with your opponent. In contrast, the computer’s goals are simple and rigid; the computer deals only with the outcome of the chess game.
Pure AI. We need to draw a distinction between “pure AI” and computer simula- tion. Pure AI is an approach that seeks to accomplish a task as efficiently as possible,
even if the computer’s processes are completely different from the processes used by humans. For example, the most high-powered computer programs for chess will evalu- ate as many potential moves as possible in as little time as possible (Michie, 2004). Chess is an extremely complex game, in which both players together can make about 10128possible different moves—more than the total number of atoms contained in our universe. Consider a computer chess program named “Hydra.” The top chess players in the world make a slight error about every ten moves. Hydra can identify this error— even though chess experts cannot—and it therefore wins the game (Mueller, 2005).
Researchers have designed pure AI systems that can play chess, speak English, or diagnose an illness. However, as one researcher points out,
AI systems typically confine themselves to a narrow domain; for example, chess- playing programs don’t usually speak English. They tend to be brittle, and thus break easily near the edges of their domain, and to be utterly ignorant outside it. I wouldn’t want a chess-playing program speculating as to the cause of my chest pain. (Franklin, 1995, p. 11)
Computer Simulation. As we have seen, pure AI tries to achieve the best possi- ble performance. In contrast, computer simulation or computer modeling attempts to take human limitations into account. The goal of computer simulation is to design a system that resembles the way humans actually perform a specific cogni- tive task. A computer simulation must produce the same number of errors—as well as correct responses—that a human produces (Carpenter & Just, 1999; Thagard, 2005).
Computer-simulation research has been most active in such areas as language processing, problem solving, and logical reasoning (Eysenck & Keane, 2005; Thagard, 2005). For example, Carpenter and Just (1999) created a computer-simulation model for reading sentences. This model was based on the assumption that humans have a limited capacity to process information. As a result, humans will read a difficult sec- tion of a sentence more slowly. Consider the following sentence:
The reporter that the senator attacked admitted the error.
Carpenter and Just (1999) designed their computer simulation so that it took into account the relevant linguistic information contained in sentences like this one. The model predicted that processing speed should be fast for the words at the beginning and the end of the sentence. However, the processing should be slow for the two verbs, attacked and admitted. In fact, Carpenter and Just demonstrated that the human data matched the computer simulation quite accurately.
Interestingly, some tasks that humans accomplish quite easily seem to defy com- puter simulation. For example, a 10-year-old girl can search a messy bedroom for her watch, find it in her sweatshirt pocket, read the pattern on the face of the watch, and then announce the time. However, a computer cannot yet simulate this task. Computers also cannot match humans’ sophistication in learning language, identify- ing objects in everyday scenes, or solving problems creatively (Jackendoff, 1997; Sobel, 2001).
Current Issues in Cognitive Psychology 19 The Parallel Distributed Processing Approach. In 1986, James McClelland,
David Rumelhart, and their colleagues at the University of California, San Diego, published an influential two-volume book called Parallel Distributed Processing (McClelland & Rumelhart, 1986; Rumelhart et al., 1986). This approach contrasted sharply with the traditional information-processing approach. As we discussed on pages 9 to 11, the information-processing approach emphasizes that a mental process can be represented as information progressing through the system in a series of stages, one step at a time.
In contrast, the parallel distributed processing (or PDP) approach argues that cognitive processes can be understood in terms of networks that link together neuron- like units; in addition, many operations can proceed simultaneously—rather than one step at a time (Fuster, 2003; O’Reilly & Munakata, 2000). Two other names that are often used interchangeably with the PDP approach are connectionism and neural networks.
The PDP approach grew out of developments in both neuroscience and artificial intelligence—the two topics we have just discussed. During the 1970s, neuroscientists developed research techniques that could explore the structure of the cerebral cortex, the outer layer of the brain that is responsible for cognitive processes. One important discovery was the numerous connections among neurons, a pattern that resembles many elaborate networks (Rolls, 2004; Thagard, 2005).
This network pattern suggests that an item stored in your brain cannot be local- ized in a specific pinpoint-sized region of your cortex (Fuster, 2003; Woll, 2002). Instead, the neural activity for that item seems to be distributed throughout a section of the brain. For example, we cannot pinpoint one small portion of your brain in which the name of your cognitive psychology professor is stored. Instead, that infor- mation is probably distributed throughout thousands of neurons in a region of your cerebral cortex.
The researchers who developed the PDP approach proposed a model that simu- lates many important features of the brain (Levine, 2002; Woll, 2002). Naturally, the model captures only a fraction of the brain’s complexity. However, like the brain, the model includes simplified neuron-like units, numerous interconnections, and neural activity distributed throughout the system.
During the time that some researchers were learning about features of the human brain, other researchers were discovering the limitations of the classical artificial intel- ligence approach. This classical approach viewed processing as a series of separate operations; in other words, processing would be serial. During serial processing, the system handles only one item at a given time; furthermore, the system must com- plete one step before it can proceed to the next step in the flowchart.
This one-step-at-a-time approach may capture the leisurely series of operations you conduct when you are thinking about every step in the process. For example, a classical AI model would be appropriate when you are solving a long-division prob- lem (Leahey, 2003).
In contrast, it is difficult to use classical AI models to explain the kinds of cogni- tive tasks that humans do very quickly, accurately, and without conscious thought. For example, these AI models cannot explain how you can instantly perceive a visual
scene (Leahey, 2003). Glance up from your book, and then immediately return to this paragraph. When you looked at this visual scene, your retina presented about one million signals to your cortex—all at the same time. If your visual system had used serial processing in order to interpret these one million signals, you would still be processing that visual scene, rather than reading this sentence! Many cognitive activi- ties seem to use parallel processing, with many signals handled at the same time, rather than serial processing. On these tasks, processing seems to be both parallel and distributed, explaining the name parallel distributed processing approach.
Many psychologists welcomed the PDP approach as a groundbreaking new framework. They have developed models in areas as unrelated to one another as reading disabilities (Welbourne & Ralph, 2007), consciousness (Kashima et al., 2007), and interpersonal attachment (Fraley, 2007). Researchers continue to explore whether the PDP approach can adequately account for the broad range of skills represented by our cognitive processes.
Keep in mind that the PDP approach uses the human brain—rather than the serial-computer—as the basic model (Woll, 2002). This more sophisticated design allows the PDP approach to achieve greater complexity, flexibility, and accuracy as it attempts to account for human cognitive processes.