Age effects on category learning, categorical perception, and generalization

Abstract

Age deficits in memory for individual episodes are well established. Less is known about how age affects another key memory function: the ability to form new conceptual knowledge. Here we studied age differences in concept formation in a category-learning paradigm with face-blend stimuli, using several metrics: direct learning of category members presented during training, generalization of category labels to new examples, and shifts in perceived similarity between category members that often follow category learning. We found that older adults were impaired in direct learning of training examples, but that there was no significant age deficit in generalization once we accounted for the deficit in direct learning. We also found that category learning affected the perceived similarity between members of the same versus opposing categories, and age did not significantly moderate this effect. Lastly, we compared traditional category learning to categorization after a learning task in which a category label (shared last name) was presented alongside stimulus-specific information (unique first names that individuated category members). We found that simultaneously learning stimulus-specific and category information resulted in decreased category learning, and that this decrement was apparent in both age groups.

Healthy memory function involves both the ability to remember details of past events (memory specificity), and the ability to link across related events to build new knowledge (memory generalization). Prior research has established that advanced age reduces memory specificity across a wide variety of metrics (Brainerd & Reyna, 2015; Hashtroudi et al., 1989; Old & Naveh-Benjamin, 2008; Yassa et al., 2011) but less is known about how age affects the ability to acquire new general knowledge and apply it in novel situations.

Aging studies of categorization have often shown age deficits in direct learning of category labels during a training phase (T. Davis et al., 2012; Filoteo & Maddox, 2004; Glass et al., 2012; Hess & Slaughter, 1986a; Mata et al., 2012; Racine et al., 2006), but not for every type of category structure or learning task (Filoteo & Maddox, 2004; Glass et al., 2012; Hess, 1982; Hess et al., 1996). Further, the body of literature devoted to category learning in older adults is much smaller than that of other types of memory deficits, and a clear consensus about the underlying cause of age deficits in category learning has not emerged. Some have suggested that older adults use less effective strategies to group items (Maddox et al., 1998, 2010), that deficits arise when the complexity of the category structure exceeds older adults’ capacity (Rabi & Minda, 2016; Racine et al., 2006), or that difficulty encoding the individual category members negatively impacts older adults’ category-learning abilities (T. Davis et al., 2012; Hess & Slaughter, 1986b). Thus, much remains unclear about the nature of age differences in acquiring category knowledge.

In addition to mixed findings on direct learning of category labels, only a small subset of aging studies have included a generalization test to see how well older adults apply category labels to new category members not presented in a training phase, which is an important hallmark of category knowledge (H. P. Davis et al., 1998; Gouravajhala et al., 2020; Hess & Wallsten, 1987; Mata et al., 2012; Wahlheim et al., 2016). While some have found age deficits in generalization (H. P. Davis et al., 1998; Hess & Wallsten, 1987), it is possible that generalization is not impaired above-and-beyond age deficits in initial learning (Mata et al., 2012). Indeed, some models of categorization posit that categories are represented by individual category members encountered during the training phase, and that generalization involves comparing the similarity of a new item to category members stored in memory (Kruschke, 1992; Medin & Schaffer, 1978; Nosofsky, 1986). According to these models, the quality of memory for individual training items is the primary limitation on subsequent generalization. One piece of evidence for these models arises when subjects are better at categorizing old training items than comparable generalization items, suggesting that subjects are using explicit memory for item-label associations learned during training.

Given their well-known deficits in episodic memory, it follows that older adults would have difficulty relying on memorization of training items to support later generalization. Indeed, prior work suggests that older adults’ deficits in remembering specific details can hinder their ability to form abstract category representations (Hess & Slaughter, 1986a, 1986b), and older adults do not use a memorization strategy as effectively as young adults (Gouravajhala et al., 2020; Wahlheim et al., 2016). However, there is evidence that older adults’ may not have a specific deficit in memorization, but rather a general difficulty employing a variety of categorization strategies (Gouravajhala et al., 2020; Wahlheim et al., 2016). Thus, older adults may have difficulty encoding item-label associations through rote memorization, and they may have other deficits that make it difficult to compensate when memorization fails.

In addition to fostering explicit knowledge of category labels, category learning has been shown to change perception of category members, improving discrimination along dimensions relevant for categorization and affecting the perceived similarity among category members (Beale & Keil, 1995; Goldstone, 1994; Livingston et al., 1998; Pisoni et al., 1982). When asked to rate the similarity between category members after category learning, participants tend to rate members of the same category as more similar to one another than members of opposing categories (Goldstone et al., 2001; Livingston et al., 1998), even when physical similarity within and across category boundaries is equated (S. R. Ashby, Bowman, & Zeithamova, 2020). The difference in perceived similarity for members of the same category vs. opposing categories may be referred to as a category bias in perceived similarity, and this category bias closely tracks generalization accuracy in young adults (S. R. Ashby et al., 2020).

While the category bias in perceived similarity has not been investigated in older adults, there is evidence for age-related differences in the relationship between perceptual processing and categorical processing. Research in speech perception has shown that older adults have less strict categorical boundaries between speech sounds (Wang et al., 2017), which can persist above-and-beyond loss of sensory sensitivity (Mattys & Scharenborg, 2014). Further, older adults may be less likely than young adults to group items based on perceptual similarity (Denney & Lennon, 1972; Wadsworth Denney, 1974), instead sorting by function (e.g., pairing a pipe with matches rather than with a saxophone) (Pearce & Denney, 1984). However, it is not clear whether older adults fail to use perceptual similarity when it is the primary means of differentiating category members and non-members or if they simply prefer functional classification when available. Together, while these studies suggest age differences in the use of perceptual similarity, it is unclear whether new learning shifts older adults’ perceptions and whether such shifts are an indicator of the ability to generalize.

As our primary goal, we sought to better understand how advanced age affects the ability to acquire new category knowledge. We were interested in three hallmarks of category knowledge: the ability to learn category labels for category members shown during training (direct learning), the ability to apply category labels to new category instances not shown during training (generalization), and changes in perceived similarity of items within- and between-categories following category learning (category bias in perceived similarity). To introduce categories (families) of individuals who resembled one another, we used computer software to blend together “parent” face photographs (see Figure 1). We used an approach introduced by S. R. Ashby and colleagues (2020) that allowed us to dissociate the effect of physical similarity from category membership. Some pairs of faces shared a parent and were thus physically similar to one another. Some of the face pairs that shared a parent were members of the same category. Other face pairs that shared a parent belonged to different categories. We expected that older adults would show deficits in acquiring new category knowledge, and we were particularly interested in whether there was a deficit in generalizing category labels above-and-beyond any deficit in direct learning of category labels. We also asked participants to rate their perceived similarity of pairs of faces before and after category learning, which allowed us to test the degree to which older adults shift their perceptions of category members following learning and whether any such shifts relate to the quality of category knowledge when tested explicitly.

Figure 1.

Example face-blend stimuli. Parent faces on the leftmost side are designated “category relevant parents” as these parents determined family membership—Miller, Wilson, or Davis—during training and generalization. Parent faces across the top are designated “category irrelevant parents” as these parents introduced physical similarity among faces but did not determine categories. Three irrelevant parents were used for training. An additional 14 irrelevant parents were used for the creating new faces used for generalization. Parent faces were never viewed by participants, only the resulting blended faces. The face blending procedure produced pairs of faces that shared a category-relevant parent and belonged to the same family (shared parent and family; example indicated with solid dark grey box), pairs of faces that shared a category-irrelevant parent and belonged to different families (shared parent only; example indicated with dashed dark grey box). Non-adjacent pairs did not share a parent or a family name (example indicated with light grey boxes).

As a secondary goal, we tested whether older adults spontaneously learn information necessary for later categorization when task instructions do not emphasize category learning. Category-learning tasks often train participants by asking them to label category members and then provide them with corrective feedback (e.g., F. G. Ashby, Maddox, & Bohil, 2002; Dunn, Newell, & Kalish, 2012; Maddox, Ashby, & Bohil, 2003; Yamauchi & Markman, 2000). However, research shows that young adults are also able to learn categories without these explicit cues (Aizenstein et al., 2000; S. R. Ashby et al., 2020; Love, 2002; Reber, Gitelman, Parrish, & Mesulam, 2003). The degree to which age affects category learning when category information is not explicitly emphasized is not clear, and it may be particularly relevant for learning categories in daily life: We are rarely presented with such explicit category-learning cues, but we are still required to learn new categories and concepts throughout our lives. In the present study, some participants learned face families through traditional, feedback-based category training of family names while others were instructed to remember the full name (unique first and shared family names) of each face through observational paired-associate training. Thus, in the latter case, category (family name) information was available to participants but task instructions did not explicitly ask participants to make connections from one individual to the next, and category information was not explicitly reinforced through feedback.

On the one hand, we reasoned that the task of encoding face-name associations would be a particularly demanding task for older adults (Naveh-Benjamin et al., 2004b; Old & Naveh-Benjamin, 2008), leaving them without additional resources to also spontaneously detect connections between individuals with the same family name. If this were the case, we would expect especially large age deficits when task instructions did not direct older adults toward the shared category label. On the other hand, research also suggests that older adults tend to extract the ‘gist’ of experiences rather than the individuating details (Dennis et al., 2007; Koutstaal & Schacter, 1997; Tun et al., 1998), and in some cases are more likely than young adults to encode details that are not directly relevant for the explicitly instructed task (Campbell et al., 2010; Weeks et al., 2016). Based on these findings, we would expect older adults to default to encoding information common across family members regardless of the explicitly instructed task, leading to a similar magnitude of age deficits regardless of whether category information is explicitly emphasized. To adjudicate between these possibilities, we compared the magnitude of age differences in each categorization metric of interest in terms of whether training included only family names or instead included both a family name and a unique first name for each training example.

Method

Participants

Eighty young adults (Mean age = 20.6 years, range 18–32 years) were recruited from the University of Oregon and the surrounding community and participated for course credit or financial compensation. Eighty older adults were recruited from the surrounding community and received financial compensation. Data collection was terminated after an a priori goal of 80 subjects from each age group had participated (40 subjects in each group × condition cell). While we did not have preliminary data on which to estimate the magnitude of potential age effects, we chose a moderate between-groups effect size of d = ~0.6 or ${\eta }_p^2=~.05$ for which we aimed to have ~80% power as computed with G*Power (Faul et al., 2007). Subsequent to data collection, two older adults were excluded: one for having participated in a similar study and one for having a score on the Mini-Mental State Examination that fell below the cut off (see below), leaving data from 78 older adults reported in all analyses (Mean age = 71.2 years, range 60–88 years). Detailed demographic information including the gender, race, and ethnicity make-up of each age and training group is presented in Table 1. All experimental procedures were approved by Research Compliance Services at the University of Oregon.

Table 1.

Demographics separated by age and training group

	Young adults		Older adults
	Feedback-based	Paired-associate	Feedback-based	Paired-associate
N (N female/male/non-binary)	40 (27/12/1)	40 (26/14/0)	37 (25/12/0)	41 (30/11/0)
Mean Age (range)	20.6 (18–32)	20.6 (18–27)	71.5 (60–88)	70.9 (60–88)
Ethnicity/Race
Hispanic	7 (17.5%)	6 (15%)	1 (2.7%)	0 (0%)
American Indian	1 (2.5%)	0 (0%)	0 (0%)	0 (0%)
White	2 (5%)	2 (5%)	1 (2.7%)	0 (0%)
More than 1 race	2 (5%)	2 (5%)	0 (0%)	0 (0%)
Unreported	2 (5%)	2 (5%)	0 (0%)	0 (0%)
Non-Hispanic	32 (80%)	34 (85%)	34 (92.0%)	38 (92.7%)
White	27 (67.5%)	27 (67.5%)	32 (86.5%)	37 (90.2%)
Asian	2 (5%)	2 (5%)	0 (0%)	0 (0%)
Black	1 (2.5%)	2 (5%)	2 (5.4%)	1 (2.4%)
More than 1 race	1 (2.5%)	3 (7.5%)	0 (0%)	0 (0%)
Unreported	1 (2.5%)	0 (0%)	2 (5.4%)	3 (7.3%)
White	1 (2.5%)	0 (0%)	1 (2.7%)	1 (2.4%)
Unreported	0 (0%)	0 (0%)	1 (2.7%)	2 (4.9%)

Percentages in parentheses reflect the percent of the age × training group that self-reported as belonging to a given ethnicity/race.

As part of an initial session on a separate day, participants in both age groups completed a neuropsychological battery that included the Mini-Mental State Examination (MMSE; Folstein, Robins, & Helzer, 1983) and portions of the Wechsler Adult Intelligence Scale IV (Wechsler, 1997). Only older individuals with a score of 25 or better on the Mini-Mental State Examination were included in the present study. Neuropsychological testing scores from one young adult subject were missing, but the subject was retained for the experimental task. Scores from the neuropsychological tests separated by age and training condition are presented in Table 2.

Table 2.

Neuropsychological test scores separated by age and training group

	Young adults				Older adults
	Feedback-based		Paired-associate		Feedback-based		Paired-associate
	Mean (SD)	Range	Mean (SD)	Range	Mean (SD)	Range	Mean (SD)	Range
Mini-Mental State	29.2 (1.1)	26–30	29.3 (1.0)	26–30	29.1 (1.2)	25–30	29.0 (1.2)	26–30
WAIS-IV: IQ	104.3 (10.6)	85–130	107.3 (12.4)	86–147	115.9 (13.1)	89–143	116.7 (12.5)	92–141
Digit span	27.9 (4.4)	19–40	29.3 (4.9)	22–46	27.4 (5.3)	17–37	28.5 (5.7)	26–30
Arithmetic	14.4 (3.1)	8–19	15.6 (3.1)	8–22	15.8 (3.7)	8–21	14.7 (3.1)	8–20
Matrix reasoning	17.6 (4.6)	8–25	19.2 (4.0)	10–24	13.7 (4.5)	6–22	14.6 (4.6)	6–24
Visual puzzles	16.8 (4.3)	7–23	16.7 (4.8)	7–26	13.7 (4.4)	7–24	14.5 (4.9)	7–26
Vocabulary	40.9 (6.8)	22–53	41.2 (7.2)	17–55	48.2 (6.4)	27–57	48.4 (5.8)	29–57
Information	16.3 (4.0)	8–24	16.2 (4.6)	7–23	19.8 (3.8)	10–26	18.6 (4.6)	9–26
Symbol search	34.5 (7.9)	21–56	34.4 (8.7)	14–57	27.7 (7.3)	17–48	30.2 (7.0)	18–44
Coding	71.6 (14.9)	50–106	76.5 (13.2)	43–103	63.2 (13.9)	37–93	64.7 (14.6)	30–107

WAIS subtests are given in terms of raw scores

In order to characterize our samples in terms of overall cognitive abilities, we computed separate age group × training condition ANOVAs on each neuropsychological measure listed in Table 2. There were no differences between groups in terms of MMSE scores (all F’s < 1.9, p’s > .17, ${\eta }_p^2<.02$). For WAIS measures, there were no significant main effects of training condition (all F’s < 3.1, p’s > .08, ${\eta }_p^2<.021$), but several significant main effects of age group. The Full-Scale IQ that aggregates over the individual indices was higher in the older adult sample (M = 116.3, SD = 12.7) compared to the young adult sample (M = 105.8, SD = 11.5). This score is a comparison to an age-matched normative sample with a mean of 100 and a standard deviation of 15. This means that the present sample of older adults is, on average, just over 1 standard deviation above their age-matched peers whereas our young adult sample was, on average, closer to their population average. To ensure that any between-group differences were not fully accounted for by differences in overall cognitive abilities, we included the Full-Scale IQ measure from the WAIS as a covariate in all analyses.

Examining age differences in the raw scores on individual WAIS measures can help unpack the age differences in overall IQ. Young adults had better performance than older adults on measures of processing speed (symbol search F(1,153) = 20.07, p < .001, ${\eta }_p^2=.12$; coding F(1,153) = 19.77, p < .001, ${\eta }_p^2=.11$) and perceptual reasoning (matrix reasoning F(1,153) = 35.67, p < .001, ${\eta }_p^2=.19$; visual puzzles F(1,153) = 12.78, p < .001, ${\eta }_p^2=.08$). Older adults had better performance than young adults on measures of verbal comprehension (vocabulary F(1,153) = 48.33, p < .001, ${\eta }_p^2=.24$; information F(1,153) = 19.46, p < .001, ${\eta }_p^2=.11$). There were no significant main effects of age for working memory measures (digit span F(1,153) = 0.60, p = .44, ${\eta }_p^2=.004$; arithmetic F(1,153) = 0.26, p = .61, ${\eta }_p^2=.002$). There was, however, a significant age × training group interaction effect for the arithmetic subtest [F(1,153) = 4.87, p = .029, ${\eta }_p^2=.03$], which was driven by numerically higher scores for older compared to young adults in the feedback group, but numerically higher scores for young compared to older adults in the paired-associates group. No other interaction effect reached significance (F’s < 1.1, p’s > .31). Overall, these effects were largely in line with typical findings in aging: age deficits in measures of fluid intelligence and age advantages in measures of crystallized intelligence. Given the overall difference in IQ, it seems that the older adult advantage in measures of verbal comprehension were large enough to offset their deficits in processing speed and perceptual reasoning.

Materials

Stimuli were images of grayscale blended faces constructed by morphing two unaltered face images together using FantaMorph Version 5 by Abrosoft. Faces were drawn from the FACE database (Ebner et al., 2010), the Dallas Face Database (Minear & Park, 2004), the Computer Vision Laboratory Database (Peer, 1999), and an internet image search. It was important to our experimental design that the morphed faces look realistic and that pairs of face-blends could be equated for physical similarity between all blends, as we sought to measure the effects of category learning on perception above-and-beyond physical similarity. Race, gender, and age are their own preexisting categories, and the stimulus blending process is not fully automated so as to allow quick generation of multiple stimulus sets. We thus chose a single race, gender, and age group to serve as the parent stimuli in the present study. Prior work has shown that white Americans exhibit greater own-race effects in face recognition than non-white Americans (Herzmann et al., 2011; Meissner & Brigham, 2001), and we expected our sample to be majority white based on U.S. population demographics, especially demographics of our older age group. It was thus most sensible to use white faces for the stimulus set. Because of greater variability in female hair styles that can make the blending process difficult and because females show less own-gender bias than males (Mishra et al., 2019), we selected male faces for the stimulus set. Lastly, because young adults tend to show greater own-age bias, and older adults do not always show an own-age bias (He et al., 2011; Wiese et al., 2008), we selected relatively young faces for the stimulus set. An example set of stimuli, including the parent faces, is presented in Figure 1. While we took steps to reduce the effects of own-group vs. other-group effects in our stimuli, we also evaluated their role where possible. We conducted analyses directly comparing participants based on race and gender for each of our measures of interest, as well as conducting control analyses to ensure that these factors did not drive our main findings. However, it was not possible to address the issue of own-age bias analytically separate from our interest in age deficits. We return to this limitation in the discussion section.

Training set

Figure 1 displays an example training set and how it was generated. The selection of category relevant and category irrelevant parent faces was random and differed across participants. For each participant, three faces were randomly selected (from a total set of twenty possible parent faces) to serve as the category-relevant parent faces that determined category membership (Miller, Wilson, Davis). An additional three faces were selected to serve as category-irrelevant parent faces in creating the training set. Each of the three category-relevant parent faces was individually morphed with each of the three category- irrelevant parent faces with equal weight given to each parent face (50/50 blend). The resulting nine blended faces were then used as training stimuli during the training task described below. Faces blended with the same category-relevant parent shared perceptual similarities and belonged to the same family category (denoted by the same last name during the training phase) while faces blended with the same category-irrelevant parent shared perceptual similarities but belonged to different families. Thus, the blending process produced face stimuli that shared physical traits both within and across category (family) boundaries and physical similarity alone was not diagnostic of category membership. Although the structure of the stimuli was the same for all participants, the training faces and the family categories were different for each participant as the selection of category relevant and irrelevant parent faces was random. Thus, any item effects, such as greater or weaker similarity among blends of a particular parent face, were unlikely to drive any effects of interests.

Categorization test set

The categorization test set consisted of 51 face-blend stimuli. In addition to the 9 training faces (old items), the test set included 42 new stimuli created by blending the 3 category-relevant parent faces with the remaining 14 parent faces not used during training. The 42 new stimuli served as generalization items during the categorization test.

Experimental procedure

All participants completed the experimental phases in the following order: initial exposure of the training stimuli, pre-training ratings of similarity between training stimuli, training (feedback-based or paired-associate), post-training ratings of similarity between training stimuli, cued recall of names for training stimuli, old/new recognition of faces, and family name generalization (Figure 2). The order of tasks was the same for all participants because completing some tasks before others could potentially contaminate results. For example, completing the categorization test before recognition would contaminate the old/new recognition results since the same new faces were used both for category generalization and as recognition lures. Participants were given instructions for each task immediately before completing it and were not told about the nature of any task beyond the current task.

Figure 2.

Task procedure. Participants in all groups first passively viewed all nine training stimuli once prior to any experimental task. All participants also rated the similarity of all possible pairs of faces before and after the training phase. Participants were randomly assigned to complete either feedback-based or paired-associate training. In feedback-based training, participants were presented with a face, responded which family they thought the face belonged to, were told whether they were correct or wrong, and told the correct family label. In paired-associate training, participants were shown a face along with its full name and made a prospective memory judgment (PMJ). Following training and post-learning similarity ratings, all participants completed a cued-recall phase. Those in the paired-associate condition were shown a face and were first asked to recall that person’s first name followed by their family name. Those in the feedback-based condition were only asked to recall the family name, as they were not provided with first name information at any point in the task. Next, all participants completed an old/new recognition task (results not reported here). Lastly, participants completed a surprise categorization task using all faces from the recognition phase and categorized them into the three families.

Initial exposure

Participants passively viewed each training stimulus once prior to completing similarity ratings to give participants an idea of the overall range of similarity between items. Each stimulus was presented for 3 seconds followed by a 1 second inter-trial fixation interval, and the order of the stimuli was randomized for each subject.

Pre-training similarity ratings

Participants viewed all 36 possible pairs of the 9 training faces and were asked to rate how similar they looked on a 1–6 scale. Each face pair remained on the screen for 5 seconds and participants made their response during that time. A rating of 1 indicated that the faces were very dissimilar and a rating of 6 indicated that the faces were very similar. Given the face-blending procedure for creating the training faces (Figure 1), there were 9 pairs of training faces that shared a parent face that would later be relevant for determining family membership (shared parent, same family name), 9 pairs of training faces that shared a parent face that was irrelevant for determining family membership (shared parent, different family name), and 18 pairs of training faces that did not share any parent face (not related). Participants rated every face pair that did not share a parent face once and every pair that shared a parent face (relevant or irrelevant) twice. We repeated the shared parent pairs in order to increase the reliability of similarity estimates when computing averages across conditions because these were the critical trials comprising the category bias metric. We note that because each blended face shared a parent with some blends and did not share a parent with other blends, each individual face would appear across all three comparison conditions, depending on what face it was compared to for a given trial. It was not the case that some individual blends were shown more frequently than others. Further, the face pairings that were included in the category bias metric of primary interest (shared parent and family name vs. shared parent only) were all shown twice during each rating phase, equating these two conditions in terms of exposure to individual face pairings.

The order of pair presentation and the left-right position of the faces was randomized for each subject, with the constraint that repeated pairs were presented once with each left-right positioning of the faces. Prior to beginning the ratings, participants completed a practice using faces that did not appear in any other part of the task (unblended Caucasian female faces). There were six practice trials, and participants repeated the practice session until they made responses on at least 5 of the 6 practice trials.

Training

Participants were randomly assigned to either the feedback-based training or paired-associate training group. For both training conditions, there were 4 blocks of training with 4 repetitions of each training stimulus in each run, totaling 36 trials per run and 16 exposures to each face over the course of training. The order of stimuli was pseudo-randomized for each subject so that each stimulus was presented before any repeated, there were no immediate repeats of the same face, and no more than 3 faces from the same category were presented consecutively.

Feedback-based category training.

About half of participants were assigned to complete feedback-based training of family membership. During each trial, participants viewed an individual face on the screen for 3 seconds. After 3 seconds, the face was removed from the screen, the response options for family membership (Miller, Wilson, Davis) appeared on the screen, and participants made a self-paced judgment. Participants then received feedback as to whether their choice was correct or wrong and the correct family name was displayed for 1.5 seconds (e.g., ‘Correct’, ‘He is a MILLER.’) followed by a 0.5 second inter-trial fixation cross.

Paired-associate category training.

About half of participants were assigned to complete observational paired-associate training of full names, including a unique first name for each face (Brad, Andy, John, Paul, Kyle, Ryan, Tyler, Eric, or Steve) and one of the three family names (Miller, Wilson, or Davis). Family name assignment to faces was identical to the feedback-based category learning condition. Assignment of first names to faces was randomized for each participant. Participants were instructed to try to remember each face with its full name, but they were not told how they would be tested on that information. During each trial, participants viewed an individual face with its full name on the screen for 3 seconds. After 3 seconds, the face was removed from the screen, and participants made a self-paced judgment of how likely they were to remember that face-name pair on a scale from 1–4. A response of 1 indicated that they would forget the face-name pair, 2 indicated that they would probably forget the face-name pair, 3 that they would probably remember the face-name pair, and 4 that they would remember the face-name pair. Each trial ended with a 0.5 second inter-trial fixation cross. The prospective memory judgments were collected to ensure participants’ attention and to match trial structure with the feedback-based group in requiring a motor response following the presentation of the face.

Post-training similarity ratings

Post-training ratings followed the same procedures as pre-training ratings except that there was no practice phase and stimuli were presented in a new random order. The category bias in perceived similarity, our measure of learning-related shifts in perception, was defined as the difference in similarity ratings between pairs of faces that shared a parent and a family name compared to pairs of faces that shared a parent but had different family names. The pre-post design for the similarity ratings allowed us to test for a category bias in perception while accounting for any pre-existing biases or similarity differences among stimuli.

Cued name recall test

We measured how well participants were able to remember name labels learned during training. Participants viewed individual faces from training and were asked to recall the name associated with that individual. On each trial, participants in the feedback-based category-training group saw a face with the prompt ‘Last name?’ at the top of the screen. Participants were asked to type the family name and press the return key when finished. They were not reminded of the three possible family names. Participants in the paired-associate category training group underwent a similar procedure except that, for each face, there was first a prompt ‘First name?’ where participants were to type the person’s first name (also without a reminder of the 9 possible first names) followed by the ‘Last name?’ prompt. The prompt only advanced from ‘First name?’ to ‘Last name?’ when the participant pressed the return key. The first name recall was included in the paired-associate condition to further emphasize the focus on specific face information and to check for learning of the first names. Thus, any differences between training conditions for recall of last names and categorization may be affected to some extent by the inclusion of the first name recall, in addition to differences due to the training itself. In both groups, participants recalled names once for each of the 9 training stimuli and were asked to make a guess even if they were not sure of the answer. The order of faces during recall was randomized across participants.

Recognition test

In addition to the measures of category knowledge that were of primary interest, we included an old/new recognition test of training faces and new generalization faces (same as those in categorization) because we were interested in comparing measures of memory specificity and categorization for the same stimuli. However, those comparisons are not directly relevant for the current research questions of age differences in category learning and generalization. The recognition test is thus not discussed further.

Categorization test

Participants viewed all test stimuli again (same stimuli as in the recognition phase), but their task was to guess which family (Miller, Wilson, Davis) each face belonged to by pressing one of three keys on the keyboard. Each of the 9 training and 42 generalization test faces was presented once and responses were self-paced. Categorization accuracy of old (training) items was used as a measure of direct learning because the paired-associate training did not include a measure of category learning during the training phase. Categorization accuracy for the new stimuli not seen during training was used to measure the degree of category generalization.

Statistical analyses

For each metric of interest, we computed an ANOVA that included age group and training group as between-subject factors along with any relevant within-subject factor (described below). A Greenhouse-Geisser correction was applied (denoted with GG) whenever the assumption of sphericity was violated.

Age effects were of particular interest, but not all metrics ultimately showed strong evidence against the null hypothesis (no age differences). We provide Bayes Factors that indicate how strong the evidence is for the null hypothesis (no age difference) relative to the alternative hypothesis (age difference) for our metrics of primary interest (accuracy in the categorization test for old and new items, post-training category bias in similarity ratings). This allowed us to resolve whether the null finding should be interpreted as inconclusive or whether there is evidence for comparable effects across age groups. We used the Bayesian Independent-Sample Inference test (i.e., a Bayesian independent samples t-test) in SPSS to compare age groups irrespective of training condition and report the resulting Bayes Factor (BF) for each comparison regardless of its statistical significance as determined by typical p-value based inference criteria. BF₀₁ (evidence in favor of the null hypothesis) or BF₁₀ (evidence in favor of the alternative hypothesis) is reported based on which of the two was > 1 (equal evidence for null and alternative).

Results

Training

Because performance metrics collected during training were very different between the two training conditions (categorization accuracy vs. prospective memory judgments), the analysis of training data focused on age comparisons within each condition separately.

Feedback-based category training

We used the accuracy during feedback-based training to test whether those in each age group were able to learn the face families and whether older adults showed any age deficits in the ability to acquire category knowledge through explicit instruction. Although young and older adults sometimes have different response biases (Criss et al., 2014; Huh et al., 2006; Kapucu et al., 2008), we used a simple accuracy measure as is common in N-alternative forced choice tests rather than a measure that would account for bias. The equal number of items in each category ensures a bias does not strongly affect accuracy. For example, a bias toward the ‘Miller’ label would not lead to better or worse accuracy on average than a bias to use the ‘Wilson’ or ‘Davis’ label, nor would it provide any benefit/disadvantage compared to a participant picking a random label when unsure of the correct response. Nonetheless, we tested for potential age differences in response biases by computing accuracies for each category and age group separately. We found no accuracy differences across categories and no interaction between age and category (both F’s < 1.1, p’s > .35, ${\eta }_p^2’\text{s}<.02$).

Mean accuracy for each training block is depicted in Figure 3A. Categorization accuracy was above chance (.33 for 3 categories) for all blocks for both age groups (all t’s > 4.5, p’s < .001). To test for age differences in acquisition of category knowledge, we computed a 2 (age: young, older) × 4 (block: 1–4) mixed-factors ANOVA on mean categorization accuracy in each block. Results revealed a significant main effect of age [F(1,74) = 19.12, p < .001, ${\eta }_p^2=.21$] with higher overall accuracy in young (M = .72, SD = .19) compared to older adults (M = .54, SD = .19). There was also a reliable linear effect of training block [F(1,74) = 10.27, p = .002, ${\eta }_p^2=.12$] with increasing accuracy across blocks. Finally, there was a significant age × block interaction on this linear effect [F(1,74) = 10.58, p = .002, ${\eta }_p^2=.13$] such that young adults showed a steeper learning rate than older adults. Taken together, while older adults showed signs of learning the category information, they did so more slowly than young adults, resulting in an age deficit that became larger over the course of the learning phase.

Figure 3.

Training and cued-recall results. A. Mean categorization accuracy by training block for learning family names in the feedback-based training condition. B. Mean rating of prospective memory by training block for learning full names in the paired-associate training condition. In A and B, young adult means are represented with a solid line. Older adult means are depicted with a dashed line. C. Proportion of family names, first names, and full names recalled in the cued-recall test. Dark grey bars depict means for the feedback-based training group, and light grey bars depict means for the paired-associate training group. Solid bars depict means from young adults, and striped bars depict means from older adults. For A-C, errors bars depict the standard error of the mean across subjects.

Paired-associate category training

The prospective memory judgments employed during paired-associate training were not an objective measure of learning and were primarily collected to equate the trial structure between conditions. However, we analyzed them to explore whether there was any evidence that older adults perceived themselves as learning the face-name associations, and whether their perceived likelihood of remembering was lower than young adults. Mean prospective memory judgments for each block are depicted in Figure 3B. To test for age differences in memory ratings, we computed a 2 (age: young, older) × 4 (block: 1–4) mixed-factors ANOVA on mean ratings in each block. Results revealed a significant effect of age [F(1,77) = 7.05, p = .01, ${\eta }_p^2=.08$] with higher ratings in young (M = 2.96, SD = 0.52) compared to older adults (M = 2.63, SD = 0.52). The linear effect of block did not reach significance [F(1,77) = 2.94, p = .09, ${\eta }_p^2=.04$], but there was a significant age × block interaction on the linear effect [F(1,77) = 17.82, p < .001, ${\eta }_p^2=.19$]. While both groups reported higher likelihood of subsequently remembering items later in training, young adults showed a steeper increase in ratings than older adults. Thus, although the metrics for the two types of training were very different, a similar overall pattern emerged: older adults showed evidence of learning more slowly than young adults, leading to substantial age differences by the end of training.

Cued name recall

Cued recall of names served as a measure of how well participants remembered the labels presented during training. For the paired-associates group, this served as the primary objective measure of learning prior to the categorization test. The mean proportion of correctly recalled first, family, and full names are depicted separately for each age and training group in Figure 3C.

We first tested whether each group was able to generate family labels at above-chance levels. One-sample t-tests comparing scores to 0 (no recall of family names) were above chance for all training and age groups (all t’s > 8.5, p’s < .001, d’s > 1.3). Next, we tested whether the ability to remember family names differed by age or training group by computing a 2 (age: young, older) × 2 (training: feedback, paired-associate) ANOVA. The main effect of age was significant [F(1,152) = 51.25, p < .001, ${\eta }_p^2=.25$], with poorer recall in older adults (M = .46, SD = .27) compared to young adults (M = .73, SD = .26). The age deficit in recalling family names is consistent with the deficits in labeling family members during training, with cued recall adding the need to actively generate the family name rather than pick among three alternatives. There was also a significant main effect of training group [F(1,152) = 8.99, p = .003, ${\eta }_p^2=.06$], with better family name recall following feedback-based (M = .65, SD = .27) compared to paired-associate training (M = .54, SD = .28). The better family name cued recall for feedback training is consistent with the greater ease of remembering only one name (family name for the feedback group) compared to two names (first name and family name for the paired-associates group). The age × training group interaction was not significant [F(1,152) = .10, p = .75, ${\eta }_p^2=.001$], indicating that the training condition effects were comparable across young and older adults.

The paired-associate training group additionally learned a first name for each individual. One-sample t-tests comparing first name recall scores to 0 (no recall) were above chance for both age groups (all t’s > 8.7, p’s < .001, d’s > 1.3), demonstrating that both groups were able to learn and remember at least some of the unique first names. Comparing young and older adults revealed significantly lower recall of first names in older adults compared to young adults [F(1,77) = 11.52, p = .001, ${\eta }_p^2=.13$]. A similarly large effect was found when we instead considered full name recall (both first and family name correct) [F(1,77) = 12.76, p = .001, ${\eta }_p^2=.14$]. Taken together, results show that all age and training groups were able to recall training labels, but older adults’ recall was significantly poorer than young adults.

Categorization

The categorization test assessed how well participants were able to categorize two types of items: family members learned directly during training and new family members not shown during the training phase. Accuracy for training items indexed the ability to learn and remember labels, whereas accuracy for new items indexed the ability to generalize category knowledge. Categorization accuracies for old and new items are presented separately for each age and training group in Figure 4. First, we tested whether each group showed above chance generalization performance by computing one-sample t-tests separately for each age × training group on accuracy for new items (chance = .33 for three categories). All groups showed significantly above chance generalization (all t’s > 4.8, p’s < .001, d’s > 0.75).

Figure 4.

Categorization accuracy. Mean accuracy for training items (dark grey bars) and new items (light grey bars) during the categorization phase. Results are presented separately for each age and training group. Dashed line represents chance performance (33% for three categories). Error bars depict standard error of the mean.

Next, we tested for age deficits in categorization and whether age deficits were similar for categorization of training items and generalization. To do so, we submitted categorization accuracies to a 2 (age: young, older) × 2 (training: feedback-based, paired-associate) × 2 (trial type: training item, new item) mixed-factors ANOVA. Full ANOVA results are reported in Table 3. There was a significant main effect of age, with older adults showing reduced overall categorization accuracy (M = .52, SD = .22) compared to young adults (M = .71, SD = .22). Thus, there was evidence for an overall age deficit in categorization ability, which was consistent with the age deficit identified during training and in cued recall of family names. There was also a significant main effect of training group, with better categorization following feedback-based training (M = .66, SD = .22) compared to paired-associate training (M = .57, SD = .23). There was also a main effect of trial type, with better categorization of training items (M = .63, SD = .28) compared to new generalization items (M = .60, SD = .24). The advantage for categorizing old compared to new items seemed to be driven by young adults, who showed 5.6% difference in categorization accuracy, consistent with our prior work (S. R. Ashby et al., 2020). However, the advantage for old items was minimal in older adults (.0006% difference), suggesting that memorization of old items played a more limited role in categorization decisions in older adults. However, the age × trial type interaction did not reach significance. Consistent with findings from the ANOVA, a Bayesian analysis comparing age groups showed strong evidence of age differences for categorizing old items (BF₁₀ = 179,697) as well as new items (BF₁₀ = 3,607) relative to the null hypothesis of no age differences.

Table 3.

Categorization ANOVA results

Effect	df	F	p	${\eta }_p^2$
Age*	1,152	36.25	<.001	.19
Training group*	1,152	7.64	.006	.05
Trial type*	1,152	8.13	.005	.05
Age × training group	1,152	0.10	.75	.001
Age × trial type	1,152	1.50	.22	.01
Training group × trial type	1,152	2.67	.10	.02
Age × training group × trial type	1,152	.02	.90	<.001

Statistically significant effects are marked by an asterisk. df = degrees of freedom; F = F-statistic; p = p-value; ${\eta }_p^2$ = effect size

Finally, we were interested in the degree to which age differences in generalization were explainable by differences in remembering the labels for the training items, or if instead there was a generalization deficit above and beyond differences in labeling old training items. We ran a multiple regression with generalization accuracy as the outcome and age, categorization accuracy for training items, and their interaction as predictors. The overall model was significant [F(4,152) = 97.06, p < .001, adjusted R² = .71], and accuracy for training items was a strong predictor of generalization performance [ß = .82, t(152) = 12.49, p < .001]. Critically, the effect of age was not significant [ß = −.08, t(152) = −0.63, p = .53], meaning that age did not explain additional variance once the ability to remember category labels for training items was accounted for. The interaction effect was also not significant [ß = −.02, t(152) = −0.16, p = .88], meaning that the strength of the relationship between old item accuracy and generalization accuracy did not differ across age groups. Finally, we used Bayesian regression to compare a simple model including only an intercept and categorization of old items as predictors to a model that also included age group as a predictor. Evidence favored the simpler model over the model that included age (BF₀₁ = 44.4). Thus, age deficits in training accuracy, the lack of a clear old vs. new item advantage during categorization, and the lack of an age deficit in generalization once categorization of old items was taken into account all point to the primary age deficit in categorization being limited by direct learning of category labels during training.

Similarity ratings

Participants rated the similarity of pairs of faces pre- and post-training so that we could measure the extent to which learning shared family names shifted the perceived similarity within and across categories and whether any such category bias in perception was present across age groups. We first verified that our face blending procedure produced greater perceived similarity between faces sharing a parent than those that did not share a parent by computing a 2 (age: young, older) × 2 (training: feedback, paired-associate) × 3 (pair type: shared parent and family, shared parent only, no shared parent) mixed factors ANOVA on ratings from the pre-learning phase. Full ANOVA results are reported in Table 4 with means for each group’s pre-training ratings depicted in Figure 5A. As expected, the main effect of pair type was significant. Faces sharing any parent (shared parent and family name M = 4.19, SD = 0.75; shared parent only M = 4.04, SD = 0.77) were rated as more similar than pairs without a shared parent (M= 2.97, SD = 0.84; both t’s > 20, p’s < .001, d’s > 1.3). This finding confirms that participants were sensitive to the physical similarity created by blending faces using a shared parent. Unexpectedly, faces that shared a parent and later shared a family name were rated as more similar than faces sharing a parent but not a family name [t(157) = 2.27, p = .025, d = 0.20], although the family name had not yet been presented. Because relevant parents were chosen at random for each participant, this suggests that some of the features of the relevant parent faces were by chance particularly salient and had a disproportionate effect on similarity ratings prior to learning. Thus, we took the pre-training differences in similarity ratings into account when analyzing post-training differences.

Table 4.

Pre-training similarity ratings ANOVA results

Effect	df	F	p	${\eta }_p^2$
Age	1,152	1.91	.17	.01
Training group	1,152	0.16	.69	.001
Pair type*	1.8,268.7	19.98	<.001	.12
Age × training group	1,152	0.13	.72	.001
Age × pair type*	1.8,268.7	10.44	<.001	.06
Training group × pair type	1.8,268.7	2.14	.13	.01
Age × training group × pair type	1.8,268.7	1.10	.33	.007

Statistically significant effects are marked by an asterisk. df = degrees of freedom; F = F-statistic; p = p-value; ${\eta }_p^2$ = effect size

Figure 5.

Pre- and post-training similarity ratings. A. Mean similarity ratings from the pre-training phase. B. Mean similarity ratings from the post-training phase. C. Mean change in similarity rating from the pre- to post-training phases (post-ratings minus pre-ratings). Ratings of pairs of faces sharing a parent and a family name are indicated in dark grey bars, those of pairs sharing a parent only are in medium grey, and those not sharing a parent are in light grey. D. Pre-training (darker grey) and post-training (lighter grey) category bias in similarity ratings. The category bias is computed as the difference in ratings between pairs of faces sharing both a parent and a family name compared to those sharing a parent but not a family name. In the case of the pre-training ratings, the family names are not yet known to participants. Ratings are presented separately for each age and training group. Error bars depict the standard error of the mean.

There was also a reliable age × pair type interaction effect. Follow-up comparisons revealed that young adults’ ratings showed greater sensitivity to physical similarity of faces. Young adults had higher similarity ratings than older adults for faces sharing a parent (with or without shared last name; both F’s > 4.3, p’s < .04, ${\eta }_p^2>.02$), while similarity ratings for faces that did not share a parent did not differ significantly across age groups [F(1,154) = 1.32, p = .25, ${\eta }_p^2=.008$].

We next tested how perceived similarity ratings were affected by category learning. Mean ratings for each age and training group in the post-training phase are depicted in Figure 5B, and the change in ratings across phases are depicted in Figure 5C. It was clear from the pre-training ratings that participants were sensitive to our manipulation of physical similarity. We thus focused the following analyses on our primary effect of interest: the category bias in perceived similarity, computed as the difference in similarity for faces sharing a parent and a family name compared to those sharing a parent but not a family name. This comparison controls for physical similarity by ensuring that all ratings involve pairs of faces that share a common parent. We then tested what effect sharing a family name had on perceived similarity above-and-beyond physical similarity.

The mean category bias in pre- and post-training ratings is presented for each group in Figure 5D. To test for differences across age and training groups in how category learning affects perception, we computed a 2 (age: young, older) × 2 (training: feedback, paired-associate) × 2 (phase: pre-training, post-training) mixed-factors ANOVA on the category bias in similarity ratings. Full results are reported in Table 5. We found a significant main effect of phase, indicating that the category bias was overall larger post-training (M = 0.34, SD = 1.24) compared to pre-training (M = 0.15, SD = 0.83). This demonstrates that learning resulted in shifts in perceived similarity ratings, reflecting a learning-related category bias. Although we were interested in age effects on the category bias in similarity ratings, no significant age effect emerged. A Bayesian analysis revealed moderate evidence in favor of the absence of age differences for both the pre- (BF₀₁ = 7.15) and post-training category bias (BF₀₁ = 5.23).

Table 5.

Pre- v. Post-training similarity ratings ANOVA results

Effect	df	F	p	${\eta }_p^2$
Age	1,152	0.35	.55	.002
Training group	1,152	0.006	.94	<.001
Phase*	1,152	4.25	.04	.03
Age × training group	1,152	0.48	.49	.003
Age × phase	1,152	0.05	.83	<.001
Training group × phase	1,152	3.00	.09	.02
Age × training group × phase	1,152	.58	.45	.004

Statistically significant effects are marked by an asterisk. df = degrees of freedom; F = F-statistic; p = p-value; ${\eta }_p^2$ = effect size

Relationship between category bias in perception and category generalization

Our prior work demonstrated that the post-learning category bias in perceived similarity can serve to measure category knowledge, predicting subsequent generalization (S. R. Ashby et al., 2020). Here, we sought to test whether age had any effect on the relationship between the category bias in perception and the ability to generalize. Because there was a significant category effect in similarity ratings prior to training, we included both the pre-training and post-training category bias in a model with age group and training group as categorical predictors of generalization performance. We also included interactions between each categorical variable and each similarity rating variable. Full results are presented in Table 6. The model overall explained significant variance in generalization performance [F(9,147) = 12.54, p < .001, adjusted R² = .40]. Of main interest, the post-training category bias was a significant predictor of generalization even when controlling for pre-training category bias by including it in the model. None of the interaction effects reached significance, and the age deficit in generalization remained significant once perceived similarity between items was taken into account. Thus, the degree to which participants’ similarity ratings were biased by category membership was a good predictor of later generalization irrespective of age.

Table 6.

Relationship between category bias and generalization across all age and training groups

Effect	ß	df	t	p
Age group (young = 0, older = 1)*	−0.42	147	−5.97	<.001
Training group (paired = 0, feedback = 1)	0.12	147	1.76	.08
Pre-training category bias	−.03	147	−0.19	.85
Post-training category bias*	0.41	147	2.30	.02
Age × pre-training category bias	0.01	147	0.11	.91
Age × post-training category bias	−0.11	147	−0.79	.43
Training group × pre-training category bias	0.11	147	0.82	.42
Training group × post-training category bias	0.10	147	0.70	.48

Statistically significant effects are marked by an asterisk. Coding of categorical variables indicated in parentheses. ß = standardized beta parameter; df = degrees of freedom residual; t = t-statistic; p = p-value

Effects of participant gender and race/ethnicity

All of our stimuli were images of white-appearing young males, and there are known differences in how individuals perceive and remember members of own-group vs. other-group members (Anastasi & Rhodes, 2005; Bernstein et al., 2007; Meissner & Brigham, 2001; Wright & Sladden, 2003). We thus tested the degree to which race and gender effects confounded age and training group effects on categorization and category bias measures.

For the analysis of participant gender, we compared self-reported males and females, excluding one participant who identified as non-binary because it was not a large enough sample to consider as a separate group. There were no significant main effects of participant gender in categorization or in the category bias (all F’s < 1.8, p’s > .18). There was a significant interaction with gender in categorization: the 4-way age × training condition × item type × gender interaction was significant [F(1,149) = 4.97, p = .027, ${\eta }_p^2=.032$]. This effect was driven by all groups showing numerically better categorization of training items compared to generalization items, except older males in the paired associates training who showed numerically better performance for generalization items than training items. No other interactions with gender reached significance (all F’s < 1.6, p’s > .2, ${\eta }_p^2<.02$). Thus, any potential own-gender advantage for the all-male stimuli did not seem to play a large role in the current results.

To examine the effects of participant race/ethnicity, we compared non-Hispanic white participants to participants from all other racial and ethnic groups, excluding those who did not report a race and ethnicity. There were no significant main effects of race/ethnicity (both F’s < 0.5, p’s > .5, ${\eta }_p^2<.004$). There was, however, a significant race/ethnicity × training group interaction on the category bias in similarity ratings [F(1,147) = 5.51, p = .02, ${\eta }_p^2=.04$; all other F’s < 3.5, p’s > .064]. In the paired associates training, non-Hispanic whites had a numerically larger category bias in similarity ratings (M = 0.30, SD = 0.99) compared to other races (M = −0.17, SD = 1.26). In the feedback training group, non-Hispanic whites had a numerically smaller category bias in similarity ratings (M = 0.16, SD = 0.98) compared to other races (M = 0.77, SD = 1.11). Notably, race/ethnicity did not moderate any of the effects of age (all F’s < 1.5, p’s > .23, ${\eta }_p^2<.02$). Thus, as with gender, participant race did not play a major role in the current results, particularly regarding age effects.

Discussion

In the present study, young and older adults learned novel categories (families) of face-blend stimuli and then generalized category knowledge to new family members. We found age deficits in both direct learning of items presented during training and in generalization of family names to new family members. However, poorer generalization performance in older adults was explainable by their poorer performance on old training items, with no additional deficit in generalization above-and-beyond difficulty remembering category labels for training stimuli. We also measured the extent to which acquisition of category knowledge shifted perceived similarity among category exemplars. Results showed evidence of a category bias in perception without significant age moderation, and the magnitude of the category bias was related to later generalization abilities. Lastly, we tested whether older adults had particular difficulty forming category representations when task instructions emphasized individuation of stimuli and the presence of category information was not explicit in the training instructions. Results showed poorer categorization across age groups when participants were not directed to learn categories, and this effect was comparable for young and older adults. Thus, we did not see strong evidence that older adults were disproportionately hindered by the need to spontaneously detect category information.

Age deficits in acquisition and generalization of novel categories

Our first goal was to determine how age affects acquisition of new category knowledge. We found that older adults showed poorer direct learning of category members shown during training than young adults, which was reflected in training accuracy for the feedback-based training group and in categorization test accuracy for old items across both training groups. These age deficits in categorization emerged despite the older adult sample having overall cognitive abilities (indexed through Full Scale IQ) that were above average for their age cohort while young adults were more typical of their age group. Our older sample even scored particularly well on measures of existing semantic knowledge (vocabulary, information subtests), and yet they were still impaired relative to young adults in acquiring new conceptual knowledge.

This finding is consistent with a number of prior aging studies that have shown deficits in initial learning of new categories (T. Davis et al., 2012; Filoteo & Maddox, 2004; Glass et al., 2012; Hess & Slaughter, 1986a; Mata et al., 2012; Racine et al., 2006). These age deficits have emerged across a variety of category structures, which suggests that their cause is at a basic encoding level and has broad downstream effects across many types of learning. Consistent with a general encoding deficit, the present data showed a similar age deficit regardless of whether training was feedback-based and emphasized family membership or was instead a paired-associated task that emphasized learning of individual category members. Thus, much like when older adults try to acquire new specific memories, we show that age-related encoding deficits arise when older adults try to acquire new conceptual knowledge as well.

The present study also provides novel evidence that age deficits in categorization are present for complex, holistically processed face stimuli. A great deal of work in young adults has focused on dissociating memory systems supporting categorization based on simple rules relying on individual features versus categorization based on multiple features that are not easily separable and with rules that are not usually verbalizable (for review see F. G. Ashby & Maddox, 2011). Prior work in aging has sometimes shown age deficits in accuracy for the latter information integration categories (Filoteo & Maddox, 2004; Maddox et al., 2010) and/or differences in the extent to which older adults rely on an optimal information integration strategy (Filoteo & Maddox, 2004; Maddox et al., 1998). However, even these studies have typically used simple stimuli with a small number of features. Only a small number of aging studies have used rich visual stimuli like those in the present study, but results have tended to show age deficits for such stimuli as well (Hess & Wallsten, 1987; Kornell, 2010). Further, research suggests that older adults often have deficits in face recognition despite relatively intact holistic processing of faces (Boutet & Meinhardt-Injac, 2019; Konar et al., 2013; Meinhardt-Injac et al., 2014). We extend these findings to show that older adults also have difficulty forming new face categories. This suggests that there may be a common processing deficit limiting both recognition and categorization of faces in older adults.

In addition to testing direct learning of training items, we also asked whether older adults would be able to generalize family names to new family members, as generalization is a hallmark of category knowledge, and transfer of knowledge a key goal when trying to improve older adult cognition (Basak et al., 2008; Schmiedek et al., 2010; van Muijden et al., 2012). While there was an overall age deficit in generalization, it was explainable by older adults’ difficulties with direct learning of the training items. First, if older adults had a specific deficit in generalization, one would expect older adults to have poorer performance for new generalization items compared to old items. This was not the case in the present study. Instead, young adults showed this typical old item advantage (S. R. Ashby et al., 2020; Bowman & Zeithamova, 2018; Juslin, Jones, Olsson, & Winman, 2003; Lacroix, Larochelle, & Giguère, 2005), and this advantage was numerically larger in young compared to older adults. Second, multiple regression indicated that age did not explain any additional variance in generalization scores above-and-beyond the ability to categorize the training items. These results suggest that the primary driver of age differences in generalization scores was the difficulty learning the item-label relationship to form a quality category representation during training. There was little evidence for additional age deficits associated with decision-making demands incurred by extending category labels to new examples. Although we demonstrated analytically that memory for old items accounts for age differences in generalization, it would be worthwhile to confirm this finding in future studies by matching young and older adult performance on training items prior to testing generalization. If age differences in generalization are due to differences in learning of training items, then matching training accuracy should eliminate age differences in generalization. If age differences persist, then there may be aspects of generalization itself that are impaired in older adults beyond difficulty in learning initial item-label relationships.

Our findings add to existing work showing deficits in initial learning of categories (T. Davis et al., 2012; Filoteo & Maddox, 2004; Glass et al., 2012; Hess & Slaughter, 1986a, 1986b; Mata et al., 2012; Racine et al., 2006), and provide new support for the notion that generalization is not impaired above-and-beyond age deficits in initial learning (Mata et al., 2012). Our data also indicate that memorization of individual category members contributes to categorization success in young adults, as proposed by exemplar models of categorization (Kruschke, 1992; Medin et al., 1978; Nosofsky, 1987), but this strategy may be less available to older adults. In addition to the deficit in memorization, it is also possible that older adults’ deficits extend to flexible use of a variety of categorization strategies, as recently demonstrated for category rule abstraction (Gouravajhala et al., 2020; Wahlheim et al., 2016). Nonetheless, it is clear that difficulty forming high quality category representations during learning is a key limitation in older adults’ generalization abilities.

Category bias in perception in young and older adults

In addition to an explicit test of category knowledge, we also measured whether older adults showed shifts in perceived similarity ratings following learning and how such shifts compared to those observed in young adults. First, prior to learning, similarity ratings were less sensitive to physical similarity of faces in older adults than young adults, extending prior work on the diminishing role of perceptual similarity of stimuli in older adults’ cognition (Denney & Lennon, 1972; Pearce & Denney, 1984; Wadsworth Denney, 1974). Second, across the entire sample, there was a significant increase in the category bias in perceived similarity following training above-and-beyond any biases that existed prior to learning. Faces within the same family were rated as more similar to one another than equally physically similar faces that were not members of the same family, showing learning-related shifts in how family members were perceived that were comparable across age groups. In addition, the degree of category bias following learning was a positive predictor of later generalization performance.

While such categorical perception has been repeatedly demonstrated in young adults (S. R. Ashby, Bowman, & Zeithamova, 2020; Beale & Keil, 1995; Goldstone, 1994; Liberman et al., 1957; Livingston et al., 1998; Pisoni et al., 1982), it has received little attention in aging research. Prior work has shown that older adults do not always rely on perceptual similarity when items may be grouped along other dimensions (e.g., functionality; Pearce & Denney, 1984), which has been taken to mean that older adults have difficulty using perceptual similarity to derive categories. Our findings do not fully support this hypothesis as we did not find strong age differences in use of perceptual similarity. Instead, we showed a similar overall pattern in how category learning affects young and older adults’ perceptions of category members. The lack of a strong age difference in categorical perception also further supports the notion that older adults’ primary category learning difficulty is in learning the item-label relationship. In the explicit categorization tasks, an age deficit could arise if older adults know which items belong to the same category (i.e., how items are grouped) but do not remember which label corresponds to each grouping. The similarity ratings allow participants to demonstrate their knowledge of how items are grouped without the additional burden of remembering their label. That significant age differences emerged in explicit categorization but not in categorical perception suggests that removing the need to remember labels was beneficial to older adult performance.

Modulation of age effects by training task

Our last question was whether the instructed focus during training modulated any of the observed age deficits. We thus included a novel paired-associate learning condition more akin to a typical associative memory task than a typical category learning task. This training condition required participants to spontaneously detect the category information present in the task while focusing on remembering individuals. We found decreased performance in the paired-associated compared to feedback-based condition in last name recall, categorization of old faces and generalization to new faces. However, none of the condition effects interacted with age group, indicating comparable age deficits across training groups. Thus, our results suggest that older adults suffer to a similar degree as young adults when the study context does not emphasize information that will be relevant for later decisions. This finding indicates a general cost to simultaneously learning category information and individuating information rather than an age-specific deficit.

An alternative explanation for the comparable age deficits across conditions is that older adults defaulted to encoding information that was common across family members for both training tasks and made only minimal attempts to learn first names in the paired-associates task. Face-name associative memory shows among the largest age deficits in cognition (Bastin & Linden, 2007; James et al., 2008; Naveh-Benjamin et al., 2004a), and older adults may have taken advantage of the cognitive economy afforded by linking family members to a single family label to minimize demands on associative memory in the paired associates task. This explanation is also consistent with the theory that older adults often remember the ‘gist’ of past experiences but not the differentiating details (Koutstaal & Schacter, 1997; Tun et al., 1998), which may have also served to prioritize encoding of category information regardless of the training condition. Older adults may have thus avoided disproportionate deficits in the paired associates task by treating it much like traditional category learning.

Own-group effects and limitations of the present stimulus set

We used face stimuli in part because they reflect the visual complexity of items to be categorized in the real world. However, faces also contain social information, which can lead to own-group recognition biases based on the perceiver’s identity, including own-race bias (Meissner & Brigham, 2001), own-gender bias (Wright & Sladden, 2003), and own-age bias (Anastasi & Rhodes, 2005). Because the stimulus set in the present study included only young, white male faces, there was potential for social memory biases to affect current results.

We first considered whether the race and gender composition of our sample drove our findings by examining whether age and/or training groups differed in the extent to which the stimulus set matched the participants’ identities. Age and training groups were relatively balanced in terms of gender, but not in terms of race. There was a higher proportion of non-Hispanic white participants in the older group than in the young group. Differences in race across age groups are also present in the wider United States population but were somewhat larger in the present study. Because more older adults than young adults were learning own-race faces, it is possible that the magnitude of age differences was underestimated in the current study and would be larger if the stimulus set was more racially varied and/or the age groups were better matched racially. Secondly, we sought to address the potential own-race and own-gender effects analytically. We found that the effects of participant race and gender were relatively limited. Control analyses indicated that the present findings were not fully attributable to own-race or own-gender bias. However, future research using stimuli that vary in gender and race will be necessary to fully adjudicate these issues.

By design, participant age varied perfectly across groups, meaning that we could not test for own-age effects separately from general age deficits. It is possible that the overall age deficits identified in the present study were conflated by own-age bias, because the face stimuli were closer in age to participants in the young adult groups than those in the older adult groups. That is, young adults may have benefitted from learning faces of their own age group and/or older adults could have been hindered by learning from faces of another age group. In either case, the stimuli could have led to an exaggeration of age effects. This thus represents a major limitation of the present study.

While resolving this issue will ultimately require follow-up experiments, there are several reasons to believe that age deficits in the present study were not solely driven by the use of young adult faces. First, own-age effects are not detected as consistently as own-race effects (He et al., 2011; Wiese et al., 2008). Second, the size of the overall age deficit in cued recall of last names – the recall measure for which all participants have data – was comparable in the present study (g = 0.99, 95% CI = [0.66, 1.32]) to the mean age difference in recall identified meta-analytically by Rhodes and colleagues (2019) (mean g = 0.89, 95% CI = [0.75, 1.03]), indicating that the use of younger looking face stimuli did not dramatically increase the age-deficits compared to what would be expected based on prior work. Nevertheless, the magnitude of the age effects observed with the current stimulus set should be taken with caution, and the findings should be replicated with stimuli that can definitively disentangle own-age effects from age deficits.

Conclusions

In the present study, we found reliable age deficits in category learning and generalization, with generalization deficits explainable by difficulty remembering the training item-category label relationships. Despite deficits in category learning and generalization, we showed comparable effects of category learning on perceived similarity of category members across age groups: following learning, members of the same category were rated as more similar to one another than equally physically similar members of opposing categories. Thus, although older adults had difficulty linking exemplars to their category labels, their perceptual grouping of category members showed evidence of category knowledge that was comparable to young adults. Lastly, we found that age deficits were relatively consistent regardless of whether learning categories was the instructed goal of training or if instructions instead emphasized remembering individuals. Thus, the demand to spontaneously detect overlap across experiences does not pose a disproportionate challenge for older adults.