Our experiments present evidence concerning development of a fundamental cognitive capacity: the detection of abstract relations across sets of unfamiliar stimuli (in this case, a set of looming shapes) that share no surface features in common. Our focus was on the abstract relation of identity between two elements in a series of three-element sequences (ABB, AAB, and ABA), or a repetition, and on the position of the repetition in the strings. In ABB and AAB sequences, the identical elements were adjacent, instantiating late and early repetition, respectively; ABB and AAB were discriminable via the position of the repetition within the sequence. In ABA sequences, there was a nonadjacent repetition.
The results from the 11-month-olds we observed can be summarized as follows (see also and ). Outcomes of ABB versus AAB and AAB versus ABB contrasts (Experiments 1A and 2B) tell us that they can encode and discriminate repetitions in both early and late positions. Results from the AAB versus ABA contrast (Experiment 2A) tell us they can encode an early repetition and discriminate it from a nonrepetition. We did not test an ABB versus ABA contrast, but presumably 11-month-olds would also be able to encode a late repetition and discriminate it from a nonrepetition, as 8-month-olds did (Experiment 1B). Results from the ABA versus ABB contrast (Experiment 3) suggest that they cannot encode a nonadjacent repetition, although we know they can detect both early and late repetitions during test as described previously. We do not know whether an early repetition during test (viz. ABA vs. AAB) would make this easier than the failure we observed in the ABA versus ABB contrast, although this seems unlikely. The pattern of novelty and familiarity preferences implies that learning ABB and discriminating it from AAB is relatively easy (a novelty preference) and that learning AAB and discriminating it from ABB and ABA is relatively difficult (a familiarity preference). Learning ABA appears to be harder still. It is possible that ABA as tested here fell between a novelty and a familiarity preference, implying an intermediate level of difficulty, but this would be inconsistent with the larger literature on learning nonadjacent relations (e.g., Gómez, 2002
The results from the 8-month-olds provide evidence for a more limited capacity to learn and discriminate rule-bound visual sequences. The failure of 8-month-olds to learn ABB versus AAB (Experiment 1A) cannot stem from an inability to encode a late repetition, because they succeeded with ABB versus ABA (Experiment 1B). This failure, therefore, might be a “confusion” of the late repetition during encoding (ABB) with the early repetition during test (AAB), implying furthermore that the repetition was learned during habituation but its position was not encoded (and hence ABB and AAB were not discriminated). Alternatively, it might be that the 8-month-olds acquired the rule “identity anywhere” or “there is a repetition” and saw both ABB and AAB as both being instantiations of this rule (although extracting the rule was easier from the end of the string than from the beginning). The failure with AAB versus ABA (Experiment 2A) cannot be due to the inability to recognize a nonadjacent repetition during test (ABA) because they succeeded at ABB versus ABA (Experiment 1B). Failure at ABA versus ABB (Experiment 3) could be due to the inability to encode a nonadjacent repetition (like the 11-month-olds) or the inability to recognize a repetition during test. (We did not test ABA vs. AAB with either age group, reasoning that it would be even more difficult than ABA vs. ABB.) A final consideration is the possibility that the null result for the AAB versus ABB contrast fell between a novelty and a familiarity preference (see the preceding discussion of ABA learning by 11-month-olds), reflecting an intermediate level of difficulty.
Based on these results, it appears that one abstract relation that is acquired relatively early is that of repetition, and one that is acquired soon thereafter is that of position in sequence. Infants’ failure (under tested circumstances) to discriminate ABA and ABB is consistent with prior work suggesting the difficulty for both infants (Gómez, 2002
; Gómez & Maye, 2005
), and adults (Newport & Aslin, 2004
) to detect relations of nonadjacent dependency. Although in principle one might encode ABA as, say, a pair of difference relations, a more natural encoding might require the detection of nonadjacent repetition, evidently a cognitively demanding task. Infants may have perceived ABA sequences as little more than randomly ordered different elements; hence no specific relation was learned and transferred to the test stimuli. (When, difference per se as an abstract relation may be acquired is unknown.)
Sensitivity to position information in visual event sequences is also available early in life, but evidence to date suggests that it might be delayed relative to repetition. Eight-month-old infants, for example, detected violations of the sequential order of three objects, introduced one at a time, after learning the sequence during habituation (Lewkowicz, 2004
). The three objects were dropped into the scene and a unique sound was played as each hit the floor. Order violations were specified in test stimuli either by audio information (reordering of sounds, but not objects), by-visual information (reordering of objects, but not sounds), or by both, and infants responded to all three. Four-month-old infants also detected sequence violations but only under more limited circumstances—when attention was directed toward global stimulus properties, accomplished by concealing a particularly salient local feature (the impact of objects against a surface), and when serial order was specified, multimodally (i.e., visual and auditory cues jointly). There is evidence of sensitivity to serial order in causal sequences by 6-month-olds, who show recovery of interest when viewing a reversal of a previously habituated causal event (Leslie & Keeble, 1987
), but little is known at present about how younger infants would respond to such events (Cohen & Amsel, 1998
The ability to detect simple reordering of familiar objects would seem to be a requirement for sensitivity to positions of abstract relations, and the Lewkowicz (2004)
study demonstrates that the foundations of this ability may become established by 8 months when tested with a visual preference method, and when orderings of objects and sounds learned during habituation are tested with identical stimuli. Yet in the experiments reported here, 8-month-olds did not respond to variations in position, implying that at this age, sensitivity to ordering of specific elements does not, on its own, suffice for recognizing the ordering of abstract elements. Eleven-month-olds did respond to position, discriminating ABB from AAB and vice versa, and we found an asymmetry in test display preferences depending on which of these two patterns was viewed during habituation. Infants habituated to ABB showed a novelty preference, but infants habituated to AAB showed a familiarity preference (control data revealed no inherent preference for either sequence). It seems likely that having a sequence-final repetition improved performance due to a recency effect, but it is unknown how 8-month-old infants would deal with an AAB versus ABB contrast with visual sequences; a positive result would cast doubt on this explanation. Discovering a repetition early in the sequence may be especially challenging, leading to a preference for familiarity rather than novelty at test in 11-month-olds; this explanation also accounts for the failure of younger infants to learn AAB. A familiarity preference might obtain when its match with a memory representation is not yet firmly in place, as would be the case when the representation is still being actively processed (Hunter & Ames, 1988
; Roder, Bushnell, & Sasseville, 2000
), implying an increased processing load imposed by an especially difficult abstract pattern.
Of course, many questions inevitably remain open. We have not, for example, tried to address the earliest age at which rules can be acquired, and establishing the lower bound for rule learning may require experimentation from multiple methods. It might be, for example, that performance can be facilitated with an operant paradigm, such as conjugate reinforcement, which has shown to be more sensitive than visual preference paradigms when testing sensitivity to correlated visual attributes (e.g., Bhatt & Rovee-Collier, 1994
, vs. Younger & Cohen, 1983
) and when testing serial order learning (e.g., Gulya, Rovee-Collier, Galluccio, & Wilk, 1998
, vs. Lewkowicz, 2004
). Other procedural differences across studies, such as rate of presentation and training method (fixed-trial familiarization vs. infant-controlled habituation) also merit further investigation.
Our goal, however, was not to discover the earliest possible evidence of rule learning, but rather to challenge the rule learning system, to better understand how it copes with arbitrary materials, and how it develops with age. The piecemeal learning of arbitrary rules that we observed contrasts with earlier work in two intriguing ways. First, statistical and associative learning are observable in very young infants. Kirkham et al. (2002)
, for example, reported that performance in 2-month-olds, the youngest infants tested, was as strong as that of 8-month-olds, the oldest tested, in a visual statistical learning task using stimulus elements similar to those in the experiments reported here. Similarly, even newborn infants have been shown to learn associations among stimulus features and retain them for short intervals (Slater, Quinn, Brown, & Hayes, 1999
); by 3 months, memory for such associations can last considerably longer (Bhatt & Rovee-Collier, 1994
). Second, the rulewise differences that we observed with arbitrary visual materials contrast with the uniformity of learning of different rules when infants are exposed to speech (Marcus et al., 1999
); by 7 months infants are already able to recognize and generalize a broad range of rules including even the ABA versus ABB comparison that eluded the 11-month-olds we observed. This possibility is consistent with the observation that 7-month-olds are better able to acquire rules from speech than from sequences of musical tones, timbres, and natural animal sounds (Marcus et al., 2007
), and can acquire rules in visual patterns that comprise familiar objects when they are presented simultaneously (Saffran et al., 2007
; Tyrell et al., 1991
; Tyrell et al., 1993
). Results of the experiments reported here highlight the need to better understand other aspects of cognitive development in infancy, such as the role of attention and working memory, likely involved in identifying and discriminating sequential patterns, and perhaps responsible for apparent discrepancies across findings in the literature.
Finally, to the extent that our tasks require noticing a pattern and extending it to analogous cases, our results may also have implications for the development of analogy (cf. Gentner, Holyoak, & Kokinov, 2001
; Goswami, 2001
), suggesting that it too might develop incrementally, broadening as a child’s capacity to perceive abstract relations expands.