2 Inferring Cognitive Processes

2.1 Using Symbols

Do humans differ from animals and machines? Many scholars argue that humans are special because they use mental representations, a view rooted in 17th-century philosophy. Descartes (1637/1960) argued that only humans possess a soul or consciousness, and the soul’s essence is only to think. His notion of thinking resembles modern information processing. Modern echoes of Descartes are easily found. Bronowski writes that “man is distinguished from other animals by his imaginative gifts” (1973, p. 20). Bertalanffy argues that “symbolism, if you will, is the divine spark distinguishing the poorest specimen of true man from the most perfectly adapted animal” (1967, p. 36).

Modern computers refute using symbols to grant humans special status. We live in an age in which the difference between humans and machines has become blurred. Perhaps we should consider the advantages provided by representations not only for humans but also for animals or machines.

Philosopher Karl Popper (1978) proposed that representations aid survival. Thinking permits actions to be modelled, evaluated, and discarded before being performed so that humans do not rashly undertake dangerous acts. For Popper, reasoning allows our hypotheses to die in our place. Cognitive psychology also presumes that planning plays a critical role in cognition. Cognitive psychologists hypothesize that humans perceive information, construct mental models of the perceived world, and then manipulate models to plan and evaluate potential actions. We call such processing the sense-think-act cycle.

Cognitive psychologists emphasize “thinking” in the sense-think-act cycle and treat thinking as rule-governed symbol manipulation. Earlier we considered evidence supporting the information processing assumption: information processing machines produce seemingly intelligent behaviours. What other evidence supports viewing cognition as computation, as manipulating mental representations? Chapter 2 provides several examples from cognitive psychology to illustrate how observing human behaviour supports inferences about human information processes, which we cannot observe directly.

Chapter 2 focuses on the multi-store memory model central to early cognitive psychology (Shiffrin & Atkinson, 1969; Waugh & Norman, 1965). That model explains memory as multi-stage information processing. Each stage involves a different memory store with unique properties. Properties include the symbols used to store information, how long symbols exist in the store, the capacity of the memory, how information in the memory is maintained, and how information can be transferred to another memory.

Although the multi-store memory model was important historically, many modern revisions to and elaborations of it exist (Baddeley, 1986; Conway, 1997; Eichenbaum, 2002; Tulving, 1983). If more modern memory theories exist, then why do I use older research to introduce cognitive psychology’s methods?

Cognitive psychology reacted to behaviorism (Flanagan, 1984; Gardner, 1984; Leahey, 1987; Miller, 2003). Behaviorism characterized psychology as a natural science by excluding mentalistic terms from theory and by focusing on the relationships between observable stimuli and behaviours (Watson, 1913). Cognitive psychology aimed to return mentalism to experimental psychology.

The behaviorist dominance of psychology caused problems for early cognitive psychologists, who encountered difficulties publishing research in mainstream journals because editors thought that the research was too mentalistic (Mandler, 2002). Furthermore, cognitivists challenged the status quo by openly rejecting all behaviorist assumptions (Bruner, 1990; Sperry, 1993). “We were not out to ‘reform’ behaviorism, but to replace it” (Bruner, 1990, p. 3). To succeed in a behaviorist environment, cognitive psychologists needed methodologies to withstand intense behaviorist scrutiny. Cognitive psychology’s research on memory, which produced the multi-store memory model, used methodologies tempered in the crucible of behaviorist criticism.

The multi-store memory model also adheres to a philosophy of science called functional analysis. Functional analysis defines cognitive psychology’s explanatory goals and opposes the philosophy of science followed by behaviorists. I will detail the properties of functional analysis in Chapter 3 by building upon the material discussed in Chapter 2.

2.2 Partial Report and Iconic Memory

The sense-think-act cycle begins with sensing. For psychology, sensing is transduction—converting something from one form into another form. In cognitive psychology, sensing transduces energy received from the world into mental representations or symbols.

George Sperling (1960) conducted an early and influential study of transduction. He wanted to determine how much information could be seen in a single brief exposure. In one study, participants saw an array of letters and numbers (Figure 2-1) for a mere 50 milliseconds (ms). After the display disappeared, participants reported as many characters as possible. Participants retrieved characters from memory because the display had vanished. Sperling called his method the whole report condition because participants attempted to report the whole character set.

A participant in the whole report condition recalled three or four characters. However, she also reported seeing other characters but could not recall them. A limitation prevented all items from being retrieved. For instance, the memory traces of the characters quickly decayed, making them unavailable for recall if participants needed a long time to report all the items.

Sperling tested this possibility with his partial report method. Participants in a partial report task experienced a display like that in Figure 2-1, again for only 50 ms. However, immediately after the stimulus disappeared, participants heard one of three possible sounds, each a signal about which characters to report. Participants reported only the characters in the display’s bottom row (low tone), or the display’s middle row (medium tone), or the display’s top row (high tone). Participants knew that their task was to report a subset of display characters but did not know which subset in advance. The tone that they heard caused them to direct attention to the appropriate subset in memory.

In a partial report task, participants again only reported three or four characters from a row. However, because participants did not know in advance which row to report, Sperling inferred that a participant’s memory held about 9 of the 12 characters. He reasoned that most of the display remained in memory because participants could retrieve three or four items from any display row. The partial report technique indicated that display memory had a larger capacity than predicted from the whole report method. Presumably, participants could report only a small number of items before the whole memory disappeared.

Sperling varied the partial report method to infer further details about the memory for brief visual displays, called iconic memory (Neisser, 1967). For example, Sperling delayed the tone signalling the to-be-reported row; a 1-second delay caused partial report accuracy to fall to the same level as that observed using the whole report technique. He concluded that items persisted in iconic memory for less than a second.

Other variations of the partial report method permitted Sperling to conclude that iconic memory represents the visual properties of characters. In a typical partial report condition, the display disappears by being replaced with a dark visual field. However, if the display is replaced with a bright visual field, then performance is poorer. The bright stimulus following the display erases or masks memory contents. Brightness masking uses a bright stimulus to erase iconic memory’s contents.

Sperling also discovered that iconic memory did not encode other display character properties. He conducted a partial report experiment that cued participants to report only the letters (or the numbers) in a display. Sperling found partial report performance to be no better than whole report performance. Iconic memory represents visual properties of characters but not abstract properties such as character type. Iconic memory illustrates a sensory register, which holds information briefly in some sensory format (e.g., visual, auditory). Other sensory registers, such as echoic memory for auditory information, also exist (Neisser, 1967). Sensory registers represent sensory information and do not encode more abstract properties, such as character type.

Sperling’s partial report method provides a pioneering example of using experimental observations to infer properties of mental representations. Sperling measured only how many characters participants could correctly report from the memory of a brief display. But he inferred many properties of iconic memory, and he described iconic memory as a high-capacity visual representation that persists for less than a second.

2.3 Primary Memory and Acoustic Confusions

To be used in the sense-think-act-cycle, information must persist for a longer duration than is offered by iconic memory. Psychologist William James wrote that “all the intellectual value for us of a state of mind depends on our after-memory of it. Only then is it combined in a system and knowingly made to contribute to a result” (1890, p. 644). He examined his own experience to argue for different kinds of memory, calling one primary memory. We consciously experience information in primary memory as part of the psychological present.

Pioneering studies in cognitive psychology investigated primary memory’s properties. One famous study by Conrad (1964) explored the format—the symbols—used to store information in primary memory. Conrad hypothesized that primary memory uses acoustic properties to represent items.

To test this hypothesis, Conrad presented sequences from a 10-letter alphabet. Five letters (B, C, P, T, and V) defined one group that sounded similar to one another when pronounced. Five other letters (F, M, N, S, and X) defined a second group with similar sounds. The two groups, however, did not sound similar to one other. Conrad predicted that, if primary memory stored items acoustically, then much confusion would occur between letters belonging to the same group, and less confusion would occur between letters belonging to different groups.

Conrad presented participants with six-letter sequences from his alphabet. Each sequence was presented visually, one letter at a time. A new letter appeared every 750 ms. Thus, Conrad used much longer display durations than those that Sperling used. As a result, Conrad was not studying iconic memory. After the sixth letter was presented, participants wrote the sequence down—from memory—in order. Conrad used recall accuracy as his dependent measure. A correctly recalled letter was written in the correct location in a six-letter sequence.

A participant made an error when she confused the correct letter with an incorrect one. Conrad summarized recall errors in a confusion matrix like that in Table 2-1. Each row in the matrix corresponds to a presented (correct) letter. Each column corresponds to an (incorrect) response to the presented letter. The numbers in a row indicate how many times participants made a particular error. For instance, the top row indicates that B was incorrectly recalled as C 13 times, incorrectly recalled as P 102 times, incorrectly recalled as T 30 times, and so on.

Conrad used results like those in Table 2-1 to note that participants confused a presented letter with a letter in its sound-alike group more frequently than with a letter in the other group. For example, consider the total number of confusions (790) between letters all belonging to the first sound-alike group. The total sums up the numbers in the grey area in the upper left of Table 2-1. In contrast, only 233 confusions occurred between letters in one group and letters in the other group (the sum of the white area in the upper right of the table). We see similar results when we compare the sums of the other two quadrants in Table 2-1.

Conrad also compared the confusion matrix from his (visual) recall study with another one from a study in which participants heard spoken letters. He added white noise to each letter’s pronunciation to make each letter hard to hear correctly. Conrad found a very high correlation between Table 2-1 and the confusion matrix for the spoken letters, providing strong evidence that the recall errors in Table 2-1 reflect letter sounds.

His study again illustrates using behavioural observations to infer cognitive processes. Conrad summarized recall errors in a confusion matrix. By examining errors, he realized that confusion was more likely between sound-alike items and not between look-alike items. Conrad used his observations to infer that primary memory represents items with a code for how an item sounds when pronounced. In short, he collected evidence to answer questions about primary memory’s basic information processing properties.

2.4 Delaying Recall from Primary Memory

Why might primary memory encode sounds when representing items? We often use rehearsal to preserve information in memory. Rehearsal involves saying items aloud to keep them in primary memory and is easier if the format of primary memory makes items easier to say aloud. In short, primary memory’s acoustic representation supports a particular process, rehearsal.

How important is rehearsal for maintaining items in primary memory? Peterson and Peterson (1959) challenged primary memory by preventing rehearsal. In a given trial, participants heard one item to remember (a consonant syllable), followed by a number. Participants then counted, aloud, backward from the number by threes until the researcher signalled them to stop. Counting out loud prevented them from rehearsing the item. When signalled, participants tried to recall the item presented at the start of the trial.

Peterson and Peterson manipulated the length of the delay—the length of time that participants counted out loud—before participants tried recalling an item. Peterson and Peterson discovered that an item’s probability of correct recall dramatically decreased as the delay increased. They used an exponential function to relate recall probability to amount of delay; the equation is provided and illustrated in Figure 2-2. The figure shows that, without rehearsal, items are likely forgotten after a few seconds of delay. Rapid forgetting without rehearsal is another primary memory property inferred from behavioural measures.

2.5 Primary Memory and Recoding

In Sections 2.3 and 2.4, I described how cognitive psychologists infer the properties of primary memory. Conrad demonstrated that primary memory stores items in an acoustic format. Peterson and Peterson discovered that items in primary memory are forgotten in seconds without rehearsal.

Another property is capacity: how many items can primary memory hold? One measurement of capacity is the span of immediate memory, the longest digit sequence that someone can recall after hearing the sequence only once (Gates, 1916). It is measured with the digit-span task, which presents digit sequences of different lengths to participants to determine the longest sequence that a participant can remember. Gates used the digit-span task to determine that humans have a span of immediate memory of between seven and eight digits.

Many later studies confirmed the results from Gates, as summarized by George Miller (1956) in his famous paper “The Magical Number Seven, Plus or Minus Two.” Miller concluded that primary memory holds only between five and nine different items—the magical number. But he famously argued that researchers define the term “item” poorly. For instance, they often vary the digit-span task by presenting different stimuli, such as letters and words. Such studies reveal that the span of immediate memory is smaller for words than for letters (Bousfield & Cowan, 1964). But is the capacity for words really smaller?

Imagine the finding that primary memory can hold seven letters but only six words. However, if each word contains at least three letters, then remembering six words means that a participant also remembers eighteen letters. Which is a more appropriate “item,” words or letters used to construct words? Miller proposed that we cope with primary memory’s limited capacity by recoding items into chunks. He argued that “chunks” are more appropriate for measuring capacity than “items.”

Recoding organizes or combines many items into a single chunk. “There are many ways to do this recoding, but probably the simplest is to group the input events, apply a new name to the group, and then remember the new name rather than the individual input events” (Miller, 1956, p. 93). For Miller, primary memory could hold about seven chunks, but each chunk could represent many individual items, extending memory capacity. He wrote that recoding is “the very lifeblood of the thought processes” (p. 95). Miller argued that we constantly use recoding, which primarily involves translating information into a verbal code. The next section illustrates recoding, teaching the reader how to recode digit sequences into words.

2.6 Example: Recoding Digits into Chunks

How might we use recoding to cope with primary memory’s limited capacity? Consider one approach to coping with the digit-span task, a mnemonic technique, or memory aid, for recoding digits into chunks represented as words. The major method, from the 1830s, recodes each digit into a consonant sound (Lorayne & Lucas, 1974). Table 2-2 provides the mapping from digits to consonant sounds. In order to use the major method, the reader must memorize the sounds associated with each digit.

Some simple memory aids help you to memorize Table 2-2. For example, the lowercase letter n possesses two feet, which is why its sound is associated with the digit 2. Similarly, the lowercase letter m possesses three feet, which is why its sound is associated with the digit 3. Typical memory aids for remembering the major method mappings are provided in Table 2-2. If you take a few minutes to use these rules to remember the mappings, then you can perform an impressive memory feat by the end of the chapter.

The major method begins by using the Table 2-2 mappings to recode to-be-recalled digits. For example, consider one 10-digit sequence used in the study by Gates: 2574638197. The major method converts the 10 digits into a consonant sound sequence: n-l-k-r-sh-m-v-t-b-k. Consult Table 2-2 to see the rules used to generate the sounds.

The major method’s second step converts sounds into chunks, where each chunk represents several digits. Vowel sounds are inserted between consonant sounds to combine multiple consonant sounds into a single word. We use vowel sounds because vowels do not represent any digits, so they can be added freely.

For instance, consider one grouping for the consonant sound sequence for 2574638197: n-l, k-r-sh, m-v, t-b-k. By adding vowel sounds, we convert the consonants into a four-word sequence: nail, crash, movie, tobacco. Each word chunks two or three consonant sounds together to represent two or three digits in a single item. Four words are easier to remember than are 10 digits.

We recall a digit sequence by reversing the major method. First, we recall a word from memory. Second, we extract the word’s consonant sounds. And third, we convert the consonant sounds into digits—our responses to the experimenter’s running a digit-span task. We repeat the process for each word held in primary memory.

Recoding and chunking using the major method become faster and easier with practice. Table 2-3 provides some additional examples of using the major method for several other digit sequences studied by Gates. The examples in Table 2-3 suggest a second recoding stage to make the major method even more efficient. In this second stage, we chunk the words in the third column into a single image to be remembered. For instance, chin and foam bring to my mind an image of a great deal of foam dripping from someone’s chin. Lake and neighbour can be combined into a single image of my neighbour Al standing in front of his dock at Hastings Lake.

2.7 Functional Dissociations of Serial Position Curves

Recoding copes with primary memory’s limited capacity. However, recoding requires additional cognitive resources. For instance, the major method requires knowing different words for chunking consonant sounds. We must also know how to map digits to sounds. We cannot store this additional knowledge in primary memory alone. Instead, we must store such general knowledge in a different, larger, and longer-lasting memory, called secondary memory by William James (1890).

Psychologists Shiffrin and Atkinson (1969) called secondary memory the long-term store and theorized, with appropriate processing, that information can be transferred to it from primary memory. For Shiffrin and Atkinson, “the long-term store is assumed to be a permanent repository of information” (p. 180).

Excellent evidence supporting the existence of different memories, primary and secondary, comes from observing serial position curves. A serial position curve plots data from a free recall experiment. In a free recall experiment, an experimenter presents participants with an item sequence to remember. At the end of the sequence, participants recall as many items as possible and in any order—hence the name free recall.

Although a free recall experiment does not constrain recall order, item order is important when researchers summarize participant performance. A serial position curve graphs an item’s recall probability as a function of the item’s position in the sequence, its serial position. Serial position curves ordinarily look like a flat-bottomed U. Figure 2-3 illustrates serial position curves for four different sequence lengths (10, 20, 30, and 40 items). A mathematical model, used to fit real experimental data (Murdock, 1962), generates each curve. The model, which predicts recall probability from an item’s serial position (x) and the list’s length (L), is also provided in the figure.

Serial position curves routinely show a higher probability for recalling the first three or four items presented in a sequence, the primacy effect (see Figure 2-3). Similarly, the last three or four items presented in a sequence also have a higher recall probability, the recency effect. The remaining (middle) items in a sequence have much poorer recall probability, producing the bottom of the U-shaped serial position curve, which increases in width as the sequence increases in length.

How do serial position curves provide evidence of the existence of both primary and secondary memory? Cognitive psychologists believe that primary memory causes the recency effect, and that secondary memory produces the primacy effect. For example, Glanzer and Cunitz (1966) incorporated the Peterson and Peterson (1959) paradigm into the free recall task. At the end of a to-be-remembered list, participants performed mental arithmetic to prevent rehearsal. Preventing rehearsal reduced the recency effect, which even disappeared with sufficiently long recall delay. Crucially, delaying recall did not alter the primacy effect.

Glanzer and Cunitz’s result supports the hypothesis that primary memory produces the recency effect. For instance, when participants can recall items in any order, they recall the most recent items first, reporting items currently being rehearsed. Preventing rehearsal means that participants cannot report such items, nor can they recall items from primary memory if a delay causes forgetting. Furthermore, finding that preventing rehearsal does not change the primacy effect indicates that the items presented the earliest are held in a store different from primary memory.

Glanzer and Cunitz (1966) reduced the primacy effect by presenting items more quickly and increased the primacy effect by repeating items. Both manipulations did not change the recency effect. Glanzer and Cunitz hypothesized that having longer times to process items, or having items repeated, aids encoding items in secondary memory. However, both independent variables do not affect the recency effect, which involves recall from primary memory, a system governed by processes (e.g., rehearsal) different from those of secondary memory.

Discovering that different manipulations affect the primacy effect and the recency effect illustrates a functional dissociation. When different factors alter a serial position curve’s different parts, we infer that the different parts reflect qualitatively different information processing. Functional dissociations permit cognitive psychologists to infer that different memory stores exist, even though the memories cannot be directly observed.

2.8 Rehearsal and the Primacy Effect

Glanzer and Cunitz (1966) found that the primacy effect diminishes with increases in the presentation rate for to-be-remembered items. What causes this decrease? Glanzer and Cunitz hypothesized that rote learning transfers information from primary memory to secondary memory. One form of rote learning involves rehearsing aloud items from primary memory, without any more detailed processing. We call such rote learning maintenance rehearsal.

Maintenance rehearsal preserves items in primary memory. Plausibly, items remaining for longer durations in primary memory have higher likelihoods of transfer to secondary memory. Longer item rehearsal increases recall probability (Mechanic, 1964; Rundus, 1971). Increasing the presentation rate reduces rehearsal, diminishing the possibility of transfer to secondary memory and decreasing recall probability.

To illustrate, consider the list of words below. If we present the list one word every 4 seconds, then participants have ample opportunity to rehearse repeatedly the first few words. Only when the middle of the list is reached will rehearsal become difficult, because participants do not have sufficient time to place all the presented items into a rehearsal loop, the set of primary memory items currently being rehearsed.

• • piano
• • snake
• • clock
• • pencil
• • lobster
• • cigar
• • star
• • house
• • pipe

In contrast, presenting the same list at a faster rate (e.g., one word every second) makes maintenance rehearsal more difficult. First, there is less time available to rehearse items. Second, the rehearsal loop will be filled faster than with a slower presentation rate. Difficulties with maintenance rehearsal decrease the likelihood that items are transferred to secondary memory and reduce the primacy effect.

Other processes of control can transfer items from primary memory to secondary memory, such as elaborative rehearsal (Craik & Lockhart, 1972). Elaborative rehearsal involves thinking about what items in primary memory mean. Such thinking requires linking new items to existing information in secondary memory. Elaborative rehearsal connects representations in primary memory to representations in secondary memory.

The list presented above permits elaborative rehearsal. Each word in the list is a concrete word used in a free recall study (Paivio & Csapo, 1969). We can easily generate mental images for concrete words. Elaborative rehearsal of these words could use secondary memory to provide a mental image for each item.

One could also enhance elaborative rehearsal by using mental images to chunk items together. For instance, we could imagine a piano played by a snake, a clock with pencil hands, a lobster smoking a cigar, and so on. Such recoding again requires knowledge already present in secondary memory.

Elaborative rehearsal involving mental imagery provides a powerful technique for improving item recall. Compare recall of concrete words (as above) with recall of more abstract words (below) (Paivio & Csapo, 1969). Participants have much more difficulty generating mental images for abstract words than for concrete words. Paivio and Csapo found, in a free recall task, significantly better memory for concrete words compared with abstract words. Elaborative rehearsal, via mental images, is easier for concrete words.

• • justice
• • ability
• • ego
• • moral
• • bravery
• • amount
• • theory
• • freedom
• • grief

The major method for recoding digits offers another example of elaborative rehearsal. To succeed, secondary memory must already hold the major method. Similarly, secondary memory must also provide the words used to chunk consonant sounds together as well as the images used to combine different words into a single chunk.

Note, too, that elaborative rehearsal will be disrupted by speeding up the presentation rate in a free recall experiment. Elaborative rehearsal will be more effective if we have more time to think about the meanings of items. Faster sequence presentations reduce the opportunity to perform elaborative rehearsal. As a result, the transfer of items from primary memory to secondary memory is impaired, reducing the primacy effect in a serial position curve.

2.9 Sentence Verification and Secondary Memory

Evidence such as the functional dissociation of the primacy and recency effects permitted researchers to infer the existence of secondary memory. Cognitive psychologists agree that secondary memory stores concept meanings. But how does secondary memory encode and organize meanings? New techniques answered such questions.

One such technique is the sentence verification task (Collins & Quillian, 1969). In that task, a participant presses one of two buttons to indicate whether sentences such as “A canary is a bird” are true or false. Researchers measure response latency or reaction time—the time from sentence presentation to button press. The sentence verification task was developed to test a particular model of secondary memory (Quillian, 1967, 1969). Quillian’s model uses a network to represent relations between concepts by encoding category membership (Figure 2-4). A network node represents each concept (the words on the left in Figure 2-4). Links between nodes represent whether one concept belongs to a higher-order category (e.g., canaries are birds). The organization of the network is hierarchical. Quillian’s model also represents concept properties by linking nodes to features. For instance, the features attached to the “Canary” node in Figure 2-4 represent that canaries can sing and are yellow.

The principle of cognitive economy governs Quillian’s network. Cognitive economy minimizes duplicate features. Rather than storing the property “Has feathers” with each node for a different bird, the network stores the property once, attached to the node “Bird.” Thus, to determine whether a canary has feathers, we find the property “Has feathers” from the “Bird” node to which “Canary” is linked.

In Quillian’s model, verifying a sentence such as “A canary breathes” involves searching the network to determine whether, starting at the “Canary” node, we can reach or retrieve the property “Breathes.” When testing the model, Collins and Quillian (1969) assumed that retrieving a property directly connected to a node takes a constant time, as does moving from one node to the next in the hierarchy. They also assumed that times are additive, so following more links to find a property takes longer. The sentence verification task tests Collins and Quillian’s predictions. In a hierarchically organized secondary memory (Figure 2-4), some sentences should be faster to verify than others.

For instance, verifying “A canary is yellow” should be faster than verifying “A canary has wings” because the first sentence only involves retrieving a property from the current node, whereas the second sentence requires retrieving a property after moving from the current node to another node in the hierarchy. Verifying “A canary has skin” should take even longer because it requires retrieving a property from a node two links above “Canary.” Collins and Quillian’s (1969) results supported the predictions. Typically, participants would take 75 ms longer to verify a sentence such as “A canary has wings” than to verify a sentence such as “A canary is yellow.” And they would take on average about 150 ms longer to verify a sentence such as “A canary has skin” than to verify the sentence “A canary is yellow.”

Collins and Quillian (1969) concluded that secondary memory has a hierarchical organization. However, they also raised new questions. For instance, participants took longer to verify false sentences than true sentences, a result that Quillian’s model could not predict. Such questions led other researchers to conduct further studies by varying the sentence verification task. New results inspired alternative secondary memory models.

For instance, category size affects sentence verification (Wilkins, 1971). Participants take less time to verify sentences involving categories with fewer members (e.g., musical instruments, sports) than sentences involving categories with more members (e.g., birds, diseases). Wilkins also found that word occurrence frequencies in language affect sentence verification speed. Sentences including higher-frequency concepts or properties take less time to verify than do sentences including lower-frequency words. Note that Quillian’s model does not represent properties such as category size and word frequency.

Similarly, the preceding stimulus affects sentence verification time. A high degree of semantic relatedness between the two sentences reduces the time for verifying the current sentence (Ashcraft, 1976), an example of a priming effect. Again, Quillian’s theory does not explain priming.

The concept or property typicality also affects sentence verification times (Ashcraft, 1978). Researchers measure typicality by having participants generate examples of a stimulus. For example, participants could generate different examples of the stimulus “Bird.” The more typical an example, the more frequently different participants will generate it. Participants generate “Robin” as an example of “Bird” more frequently than “Duck.”

Ashcraft (1978) found faster verification for sentences containing typical concepts or properties (“A robin has feathers”) than for sentences containing atypical concepts or properties (“A duck has a bill”). Note that Collins and Quillian (1969) would predict that both sentences produce the same verification time because both “Robin” and “Duck” are the same distance away from “Bird.”

When new experiments produced problematic results for Quillian’s model, the model was revised (Collins & Loftus, 1975). As revised by Collins and Loftus, the model was a semantic network (similar to that in Figure 2-4) composed of interconnected nodes. However, the semantic network abandoned the logical structure of concepts used by Collins and Quillian (1969). Instead, the semantic network of Collins and Loftus represented the semantic similarity or semantic relatedness of concepts. Two concepts with high semantic similarity share many properties. A semantic network represents semantic relatedness by having two concepts connected to many of the same nodes (the shared properties).

Spreading activation serves as the basic process in a semantic network. When nodes in the network activate, activation travels as a signal to other connected nodes, increasing activity in nodes that receive the signal. When a participant is presented with a sentence such as “A robin has feathers,” the nodes for both “Robin” and “Feathers” activate, and activation spreads from them. If activity from “Robin” and “Feathers” intersects or collides in the network, then the sentence is true.

The semantic network model addressed many limitations of the original Quillian network. For instance, the higher the semantic relationship between the two nodes, the faster the signals intersect, because the number of potential routes for activity to intersect increases with increases in semantic relatedness. The revised model also explains typicality effects because more typical concepts or properties have more links to other nodes. Priming effects can be explained by hypothesizing that network activity takes time to decay, meaning that activity produced by a previous stimulus might still exist in the network, facilitating the verification of a related sentence.

The semantic network revised Quillian’s original model. Other researchers abandoned network representations, proposing radically different secondary memory models. For example, set-theoretic models represent each concept as a collection or set of features (Rips et al., 1973). Rips et al. replace spreading activation with a process of feature comparison. To verify a sentence such as “A robin is a bird,” we compare the features for “Robin” to the features for “Bird.” If the two concepts share a sufficient number of features, then the sentence is true. Sentence verification speeds up as the number of shared features increases, which accounts for many of the results obtained from the sentence verification paradigm.

Even when researchers agree on set-theoretic representations, they dispute other details. For instance, Rips et al. (1973) presume that we represent concept categories (e.g., “Bird,” “Fish”) as feature sets, but a category member must possess particular features. For example, all birds have hearts. Requiring particular features to belong to a concept is known as the classical theory of concepts.

In contrast, prototype theory does not represent categories with definite features (Rosch, 1975; Rosch & Mervis, 1975). Prototype theory instead defines categories using the family resemblance between category members, because some members better represent a category than do others. Prototype theory represents each category member as a set of features varying in cue validity. A feature with high cue validity occurs more frequently in category members than does a feature with low cue validity. A prototype possesses many features with high cue validity. In prototype theory, we classify a new instance as belonging to a category if the new instance has high family resemblance to a category’s prototype. Family resemblance is determined by comparing the features of the instance with those of the prototype using a process sensitive not only to which features are shared but also to cue validities (Tversky, 1977).

We cannot directly observe secondary memory. We must infer its properties from other observations, such as results provided by the sentence verification task. By measuring the time taken to determine the truth of a sentence of the form “An x is a y,” and by manipulating the relationship between x and y, cognitive psychologists can infer properties of secondary memory.

However, the inferences that cognitive psychologists make must be re-evaluated constantly when new experimental results are obtained. New sentence verification results required existing models to be revised, producing new theories. Progress in cognitive psychology depends on developing more sophisticated techniques for evaluating whether one proposal (e.g., a semantic network) is more plausible than another (e.g., a set-theoretic model). I consider the logic of comparing theories in more detail in Chapter 3.

2.10 Associations, Verbal Learning, and Secondary Memory

In the previous section, I used the sentence verification task as a context for examining secondary memory. However, researchers use many other methods to study secondary memory. Results from such methods in turn support many different theories about cognitive processing. Below I briefly consider one example, associationism.

Associationism proposes links or associations between concepts; thinking of one concept causes us to think about another. Associationism originated in the writings of Aristotle in 350 BC (Sorabji, 2006). Experimental psychologists have had a long interest in associationism (Warren, 1921). Associationism also inspired a model of secondary memory called HAM, for human associative memory, which arose at the same time as the models introduced in Section 2.9 (Anderson & Bower, 1973).

A particular form of associationism inspired behaviorism. Behaviorists did not study associations between ideas; rather, they studied habits, associations between environmental stimuli and behavioural responses (Thorndike, 1932). Behaviorists aimed to make psychology a natural science by focusing only on directly observable entities (stimuli and responses) and by removing unobservable mental terms from psychological theory (Watson, 1913).

However, through the first half of the 20th century, several factors increased interest in studying an older idea, the associations between ideas. These factors included the rise of information theory, growing interest in the study of human learning and memory, and the impact of linguistics on psychology (Cofer, 1978). In studying the associations between ideas, psychologists returned to investigating unobservable mental properties, leading to experimental psychology’s verbal learning tradition. That tradition modified behaviorist methodologies to study the learning of verbal materials (Andresen, 1991; Deese, 1965; Deese & Hulse, 1967; Hunt, 1971). Although verbal learners proposed associations between mentally represented concepts, they studied objective stimulus properties (Cramer, 1968).

Stimulus meaningfulness provides an example property. Researchers measure meaningfulness by having participants generate associates of stimulus words (Noble, 1952). A word that produces many associated words has higher meaningfulness than a word that produces fewer associates. Note that cognitive psychologists define meaningfulness not by using internal semantics but by using observable behaviour. Verbal learning experiments demonstrated that words with higher meaningfulness have higher recall probability in memory experiments (Deese & Hulse, 1967).

The paired-associate learning task provided a key methodology to the verbal learning tradition. That task, invented by Mary Whiton Calkins in 1894, presents participants with pairs of unrelated words to remember (e.g., House-Tree, Robin-Dog, etc.). Typically, the paired-associate learning task uses the anticipation procedure (Pennington & Waters, 1938). In that procedure, a researcher presents the first word of a pair, and participants try to recall the second word. (On the first trial, the second word must be guessed.) After responding, participants see both members of the pair. Thus, each trial provides both a test and a learning opportunity. Researchers measure performance by counting how many times the list must be presented before a participant recalls the words perfectly. The paired-associate learning task was popular because various independent variables, such as the meaningfulness, frequency, and similarity of stimulus items, could be manipulated easily (Goss & Nodine, 1965; Underwood & Schulz, 1960). Also, the task seemed to test most directly associationist ideas of interest to verbal learners (Deese & Hulse, 1967).

The verbal learning tradition bridged waning behaviorism and rising cognitivism. Early on, verbal learners had difficulty publishing results in mainstream journals because the verbal learning approach seemed to be too mentalistic (Mandler, 2002), a problem solved when Charles Cofer founded the Journal of Verbal Learning and Verbal Behavior (Cofer, 1978; Virues-Ortega, 2006).

However, as cognitivism flourished, the verbal learning tradition became more mentalistic and emphasized principles governing secondary memory’s organization (Tulving et al., 1972). For example, HAM modelled human associative memory by forming associations between nodes representing the hierarchical structure of linguistic propositions (Anderson & Bower, 1973). We will see in Section 2.11 that verbal learners also became more cognitive because paired associate learning experiments demonstrated that unobservable properties of representations were the most powerful predictors of memory performance. The verbal learning tradition finished its conversion into modern cognitivism in 1984, the year that the Journal of Verbal Learning and Verbal Behavior changed its title to the Journal of Memory and Language.

2.11 Imagery and Secondary Memory

The previous section introduced the verbal learning tradition as well as a key method, the paired-associate learning task. The verbal learning tradition proposed that secondary memory encodes associations between concepts. Section 2.10 noted that the verbal learning tradition helped psychology to transition from behaviorism to cognitivism. The paired-associate learning task pushed verbal learning theories toward cognitivism by demonstrating that the most important predictors of memory performance could not be directly observed.

In particular, the paired-associate learning task re-established experimental psychology’s interest in another potential encoding in secondary memory, mental imagery. When we experience a mental picture, we experience a mental image. The idea that we encode concepts as mental images is as old as associationism. Aristotle believed that images represent ideas (Cummins, 1989). The first mnemonic techniques attributed to the Greek poet Simonides also used mental imagery (c. 500 BC) (Yates, 1966).

We can generate mental images for some concepts easier than we can generate images for others (Paivio, Yuille & Madigan., 1968), as we saw earlier in the two lists of words on pages 45 and 46. Such concepts are high in imagery. We measure concept imagery by having participants rate how easy or difficult it is to create a mental image for a concept (Paivio et al., 1968).

The paired-associate learning task can demonstrate that imagery predicts recall better than do traditional verbal learning variables (Paivio et al., 1968). Paivio et al. even conducted one study that controlled stimulus imageability while varying stimulus meaningfulness. The study demonstrated that meaningfulness did not affect recall performance. In fact, increasing meaningfulness decreases memory performance when imagery is controlled!

In general, Paivio’s research demonstrates that imagery is one of the best predictors of performance in memory tasks (Paivio, 1969, 1971). Paivio’s results led to another proposal about the nature of secondary memory, dual-coding theory (Paivio, 1971, 1986). According to dual-coding theory, we store concepts in secondary memory using more than one format. One is a verbal code or label. Another is a mental image. We can more easily retrieve concepts represented by both types of codes, explaining better memory of concrete concepts than of abstract concepts.

2.12 Inferring Structure, Process, and Control

In the preceding sections, I described early methods invented by cognitive psychologists to infer human information processing. The examples selected for Chapter 2 were important contributions to developing the modal memory model (Baddeley, 1990). That model is also known as the multi-store memory model, one of early cognitive psychology’s crowning achievements (Shiffrin & Atkinson, 1969; Waugh & Norman, 1965). Figure 2-5 illustrates the modal memory model’s general structure.

The modal memory model possesses several features typical of cognitive psychology. First, the model explains a very general phenomenon, memory, as an organized system of subsystems: sensory registers, primary memory, and secondary memory. Such an account illustrates cognitive psychology’s core methodology, functional analysis. I explore functional analysis in Chapter 3.

Second, the model is functional in nature. The physical natures of the component memories are not described. “Our hypotheses about the various memory stores do not require any assumptions regarding the physiological locus of these stores” (Shiffrin & Atkinson, 1969, p. 179).

Third, the model’s different components have different functions, and functions are organized in a particular fashion, with information being transferred (while being recoded) from one memory to another. For example, sensory registers such as iconic memory briefly hold a large amount of information for a short duration so that some information can be transferred to primary memory. Primary memory holds a small number of chunks, encoded acoustically, representing our experience of the present. Information in primary memory can be transferred to large-capacity secondary memory to represent concept meanings for a long duration.

Fourth, each component in the modal memory model has different structural properties. Components differ in terms of capacity, information encoding, and memory duration.

Fifth, each component of the modal memory model is governed by a different kind of process. For instance, attention can be directed to different parts of a sensory register to transmit a small amount of its contents to be transferred to primary memory. The triangular shape of the attention process in Figure 2-5 illustrates that it transfers only a limited amount of information. Similarly, information can be preserved indefinitely in primary memory via maintenance rehearsal, or it can be transferred into secondary memory via elaborative rehearsal.

Sixth, though some processes for manipulating information in the model’s stores are automatic (e.g., those causing information to be forgotten), others are under conscious control. For instance, we saw in Sperling’s partial-report technique that participants can direct attention to different parts of iconic memory. Similarly, they can choose which contents of primary memory undergo elaborative rehearsal and the nature of this elaborative processing. Maintenance rehearsal is also under our explicit control.

Importantly, the properties of the modal memory model in Figure 2-5 are all architectural. As we saw in Chapter 1, an architectural account of a computer, which describes structures, processes, and control, explains how the computer processes information. The modal memory model also provides an architectural account by detailing the structures, processes, and control of different memory stores critical to human information processing. In Chapter 3, I explore why cognitive psychologists need to identify the architecture of cognition.

2.13 How to Remember π to 100 Digits

Section 2.11 described two intersections between associationism and imagery. First, associations between ideas are associations between images. Second, imagery has an important role in memory, as revealed by Paivio’s research. We can illustrate a third intersection between associationism and mental imagery using the mnemonic technique called the method of loci (Yates, 1966), a method designed for remembering a sequence of ideas in a particular order. The method of loci stores an idea to remember in a location. The method uses familiar locations, spatially arranged in a particular order. For example, we could use the rooms of a familiar house as loci because we reach each room in a particular order by mentally “walking through the house.”

To use the method of loci to remember a sequence, we create a mental image to represent the first item to remember and “place” that item in the mental image created for the first location. We then image the second item to remember and place it in the image of the second location and so on. Note the intersection of images and associations in the method. The method of loci requires mental images but associates a new image with a familiar one by mentally linking an item’s image with a place’s image. The method also illustrates elaborative rehearsal because it links new information to already known information.

With the method of loci, recall involves mentally “walking” through the sequence of remembered locations; we perform the “walking” in a set order because locations have a particular spatial layout. At each location, we retrieve the image associated with the location earlier. As a result, we recall items in the correct order. Below I consider a concrete example of the method of loci, demonstrating use of the method to memorize long strings of digits: remembering the first 100 digits of π. We remember the 100 digits in order by combining the major method (Section 2.6) with the method of loci.

Memorizing the 100 digits begins by creating phrases for recoding and chunking the digits to remember. Table 2-4 provides one possible set of phrases. The table’s first row provides a phrase representing the first digits to remember, the next row presents a phrase representing the next digits to remember, and so on to the end of the table. We must recall the phrases in their order in the table.

I can easily generate images for the phrases in Table 2-4. I imagine the first phrase as being like a coin-operated mechanical horse ride for a child, but in this case it is a mechanical lama whose legs lope when the machine activates. I imagine the second phrase as a person wearing a Toronto Maple Leafs jersey while picking the name of a poem from a book of poetry.

Given the phrases and images in Table 2-4, we next remember the (ordered) phrases. We use the method of loci, which explains our need to create an image for each phrase. Table 2-5 provides locations that I use to memorize the images for Table 2-4 in order. As noted earlier, we must use familiar, easily imagined, and sequentially arranged locations. Rooms in a well-known house serve that purpose well. Table 2-5 lists locations in my own house. I am very familiar with them, and I encounter them in a particular order as I move through them. For these locations, I start outside the house on the front sidewalk, go up the front stairs, into the front vestibule, and so on. Of course, readers must come up with their own loci to use as memory locations in order to be able to remember π.

The procedure used to memorize the digits is straightforward. I mentally link the image created for the first phrase in Table 2-4 to the image created for the first place in Table 2-5. For example, I imagine my coin-operated “motored loping lama” operating in the middle of my front sidewalk. Next I place the second image (“leaf pick poem name”) on my front step. Then I place the third image (“fur changer mom”) in the front vestibule. I continue until I place every phrase image in one of my locations. I need practice to ensure that I remember correct phrases in the correct order; I typically practise by moving through the loci a few times. I do not need a great deal of such practice.

With surprisingly little effort, we can learn, and recall, the images in order. Memorization and practice of the major method helps to convert phrases quickly into a sequence of digits. Most of the effort comes in the first step, generating phrases for recoding digits into imageable chunks.

The mnemonic technique described above illustrates many of the information processing properties inferred by cognitive psychologists from their memory experiments. For instance, rote learning or maintenance rehearsal commits the major method to memory. Elaborative rehearsal links the major method with existing knowledge using the memory aids of Table 2-2. The major method illustrates recoding and chunking. Converting images to phrases illustrates Paivio’s dual-coding theory. Using images with the method of loci illustrates elaborative rehearsal as well as associating images together.

2.14 Chapter Summary

Cognitive psychologists hypothesize that cognition is information processing, an idea inspired by the digital computer. They assume that human thinking, like computers, is rule-governed symbol manipulation. Thus, cognitive psychologists must explain human cognition in the same way that computer scientists explain a computer’s behaviour. In Chapter 1, I presented a general approach to explaining a computer by describing its architecture. Which symbols represent information? Which processes manipulate symbols? How does the computer control the order of applying processes?

Cognitive psychologists recognize that explanations of human cognition must answer similar questions. Such psychologists aim to identify human cognition’s basic properties and do not worry about whether similarities exist between human cognition and computer information processing. Instead, cognitive psychologists assume that humans are a kind of computer and seek to determine what kind of computer humans are (Hunt, 1971).

However, cognitive psychologists face a formidable problem. Unlike computers, human participants do not permit cognitive psychologists to observe internal cognitive processes directly. As a result, these psychologists must invent new methods for collecting behavioural observations to support inferences about the properties of cognition.

In Chapter 2, I introduced many example methods: Sperling’s partial report method, confusion matrices, Peterson and Peterson’s delay of recall technique, the functional dissociation of the serial position curve, the sentence verification task, and the paired-associate learning task. All of these methods explored the properties of human memory in the early years of cognitive psychology and permitted cognitive psychologists to infer basic properties of human memory. Cognitive psychologists proposed a series of different memory stores, each defined by different properties (capacity, duration, kind of information represented, processes, and control). By the middle of the 1960s, experimental results supported one of cognitive psychology’s most influential ideas, the modal memory model.

In Chapter 3, I will show that the modal memory model provides but one example of the philosophy of science adopted by cognitive psychologists: functional analysis. I now turn to describing functional analysis and how it converts cognitive descriptions into explanatory theories.