Humans presented with word pairings in a serial order A-B seem to spontaneously form the reverse association B → A alongside the trained one A → B. This has been researched in the Paired Associate studies of the 1950s–60 s, and is evidenced by transfer effects, when learning of new word pairings that include either word A or word B is influenced by the untrained, backward B → A association (e.g., Harcum,
1953; Murdock,
1956), but also, upon presentation of word B, by direct recall of word A (e.g., Feldman & Underwood,
1957; Stoddard,
1929). Notwithstanding a debate on the equal strength of both directions of the association (Asch & Ebenholtz,
1962; Houston,
1964; for a review, see Ekstrand,
1966), such results were overall robust and authors interpreted them as human subjects readily forming bidirectional mental relations between word A and word B after a serial exposure.
From an evolutionary perspective, it matters to know whether non-human animals can also form bidirectional relations between stimuli presented serially, or if this capacity is unique to humans, and maybe related to language. Relevant information on this issue can be found in the associative learning literature, based on both Pavlovian and operant conditioning protocols.
Pavlovian conditioning protocols
One first way to investigate the bidirectionality of associations after a forward pairing A-B is the use of complex
1 Pavlovian conditioning protocols. Following
backward second-order conditioning in rats (i.e., exposure to A-US (unconditioned stimulus) then to A-B), Cole et al. (
1995, Exp. 1) have reported conditioned responding to A but also to B, the stimulus not paired with the US (A and B being 5-s auditory stimuli and the US aversive
2). Postulating that a backward untrained association B → A existed, i.e., that training with forward pairings A-B created a bidirectional association A ↔ B, could account for such results in that the chained association B → A → US would give B its new properties, but Cole et al. (
1995) interpreted their data differently. Based on the influence of varying the A-US interval, they argued instead for integration of stimulus timing across both training phases, resulting in a
temporal map (cf. Matzel et al.,
1988) representing all three stimuli, and in B becoming predictive of the US. Cole et al. (
1995) swapped training phases (i.e., A-B then A-US) in their second experiment. This procedure, called
backward sensory preconditioning, led to the same results and interpretation.
Backward sensory preconditioning was also employed by Ward-Robinson and Hall (
1996), who, though with a different theory in mind, similarly argued against an explanation in terms of a backward association B → A. Using a long auditory stimulus as A and a short visual stimulus as B, they too found conditioned responding to B in rats, but did not favor a temporal map account, as B would not predict the US given their temporal parameters. Their complementary Experiment 3 allowed these authors to propose that a representation of B, originating from the A-B pairing, was later activated by A during the A-US pairing, allowing this evoked representation to directly become associated with the US and acquire response-eliciting properties.
A similar interpretation to that in Ward-Robinson and Hall (
1996) had been given by Holland (
1981) in a rat experiment involving two forward pairings: first A-US
+ between an auditory stimulus A and an appetitive US, then A-US
− between A and an aversive US. After this, responses to the US
+ were partly inhibited. The author likewise excluded the chained association account (US
+ → A → US
−), which postulates a bidirectional association A ↔ US
+, proposing instead that A activated a representation of US
+ during the second pairing, which led to joint activation of US
+ and US
− representations, hence to devaluation of the US
+. Other rat conditioning experiments reported by Holland (
1990) support this notion that a CS (conditioned stimulus) such as A can activate the representation of a US following forward CS-US pairings. Yet, in this “representation-mediated” interpretation of conditioning, only US representations can be activated,
3 and only in a forward manner, which seems insufficient to explain the more standard,
forward sensory preconditioning (A-B then B-US), in which B would not activate any representation of A during B-US pairing, hence no A → US association would ensue; this shortfall could bring us back to assuming a temporal encoding of the stimuli.
As these three examples illustrate, there is no unified account of the directionality of associations formed in Pavlovian conditioning experiments using sequential pairings. As a result, some current Pavlovian models assume that associations are bidirectional (e.g.,
HeiDI; Honey et al.,
2020) while some do not (e.g.,
A-learning; Ghirlanda et al.,
2020).
Operant conditioning protocols
Perhaps an easier way to compare associative mechanisms between humans and non-humans is the operant procedure of Conditional Matching-to-Sample (CMTS; Cumming & Berryman,
1965), which avoids response-eliciting stimuli such as USs, and can be implemented in a reasonably comparable way across species (e.g., in humans and monkeys (Sidman et al.,
1982) or in humans and pigeons (Navarro & Wasserman,
2020)). CMTS protocols always involve a training phase followed by a test phase. In training trials, a
sample stimulus A is presented first, followed by
comparison stimuli, namely a stimulus B arbitrarily associated with A presented alongside one or several non-associated stimuli. While correct selection of B and incorrect selection of other stimuli are differentially reinforced, two or more associations are learned, A1 → B1, A2 → B2, etc., generally between visual stimuli or sometimes auditory ones. Typical test trials assess bidirectionality by direct recall – with B as sample, correct selection of A among comparison stimuli constitutes evidence that the B → A association emerged during training – and are conducted as
probes: they are non-reinforced, to exclude any learning of B → A, and interspersed within reinforced forward trials, to maintain responding.
Bidirectionality studies involving CMTS procedures are generally related to
stimulus equivalence, a notion presented by Hull (
1939) as the acquired capacity of several stimuli to evoke the same behavior in an animal, but given its current meaning by Sidman and his colleagues. According to Sidman and Tailby (
1982), stimulus relations learned in a CMTS task are conditional ones such as “If A is presented as sample, then I should answer B” (“If A, then B”), but these can furthermore be considered equivalence relations if a participant spontaneously displays three types of untrained, or
derived, relations: reflexivity (“If A, then A”), transitivity (“If A, then C,” having learned “If A, then B” and “If B, then C”), symmetry (“If B, then A,” having learned “If A, then B”). In CMTS studies on stimulus equivalence, the question of bidirectionality is thus reduced to this third property of
derived symmetry, or simply, symmetry.
It is commonly accepted that humans readily show symmetry in CMTS studies, hence non-verbal associations would also be bidirectional – but we have recently warned against a general lack of rigor in these demonstrations (Chartier & Fagot,
2022), which should caution the reader to put this consensus in perspective. In non-humans, symmetry certainly has limited empirical support, as is apparent from two thorough reviews by Lionello‐DeNolf: out of 40 studies, the author considered that 16 produced no evidence whatsoever, and she retained only seven with unambiguously positive results and no obvious alternative explanation (Lionello-DeNolf,
2009, Table 1; Lionello‐DeNolf,
2021, Table 1). Among these seven studies, two involved sea lions (Kastak, Schusterman, & Kastak,
2001; Schusterman & Kastak,
1993), five involved pigeons (Campos et al.,
2014; Frank & Wasserman,
2005; Swisher & Urcuioli,
2013,
2015; Urcuioli,
2008), and noticeably none involved primates.
Importantly, we note that all seven studies involved direct recall (probe trials), and did not limit their training procedure to A-B, but included additional pairings referred to as
identity (A-A and B-B: all of them except Campos et al.,
2014),
dual oddity (A1-A2, A2-A1, B1-B2, B2-B1: Campos et al.,
2014), or
partial symmetry (B-A for a subset of stimuli: Schusterman & Kastak,
1993), and aimed at familiarizing animals with a variable temporal position of stimuli. Thus, at least pigeons and sea lions can probably form bidirectional associations following forward pairings in such complex CMTS protocols, but it remains unknown whether they can do so spontaneously, i.e., after only A-B trials. In contrast, at least one CMTS study with mere A-B training in humans has reported derived symmetry in the very first test trials (Arntzen & Haugland,
2012). Hence, though humans appear more prone to encode bidirectional associations in CMTS procedures than non-humans, it would be crucial to know whether symmetry can emerge in non-humans after A-B training alone.
Transfer effects in conditional matching-to-sample (CMTS) studies on symmetry
It turns out that three studies using only A-B training, and the alternative testing approach of transfer effects, did find weak evidence for symmetry in pigeons (Hogan & Zentall,
1977) or capuchin monkeys (D’Amato et al.,
1985; Soares Filho et al.,
2016). The strategy used to reveal bidirectionality was to train participants with two or more A-B pairings, and subsequently have them learn reversed pairings that were either consistent (B1-A1, B2-A2, etc.), or inconsistent (e.g., B1-A2, B2-A1, etc.). Faster learning for the consistent pairings was taken as evidence that the B → A association was already present after training and transferred to the second phase, hence that forward training A-B creates bidirectional associations A ↔ B. We believe that various shortcomings may have prevented clearer demonstration of symmetry in these studies, and that transfer effects deserve further exploration.
Hogan and Zentall (
1977) separated 12 pigeons in two test groups, consistent and inconsistent, and found a difference in immediate test performance apparent in the first 28, but not 48, test trials. They found, however, no faster overall learning in the first group and concluded there was a “
minimal” (p. 13) strength of the derived B → A association. They did not run a control for stimulus effects and their design did not allow comparison between both conditions in each participant. D’Amato et al. (
1985) tested six capuchin monkeys successively with consistent and inconsistent pairings, and reported immediate test performance indicative of symmetry in two monkeys, but their test was too short (24 trials) to allow reversed pairings to be learned until criterion, and no binomial tests were applied to compare performance to chance level. Doing so (on six test trials per pairing) reveals that responses of one subject only, Dagwood, achieved statistical significance, and only in one session. Moreover, stimulus preference may have accounted for the results. Last, Soares-Filho et al. (
2016) trained one capuchin monkey on two pairings tested with consistent reversals, then on two new pairings tested with inconsistent reversals. A difference in learning length for reversed pairings suggested symmetry; however, no control was provided for the possibility that inconsistent pairings were intrinsically harder to learn than consistent ones (e.g., due to stimulus choice), nor for an effect of condition order (e.g., the second test may take longer simply due to boredom).
Finally, three further investigations of symmetry through transfer effects can be mentioned: Richards (
1988), who found a slightly faster learning for consistent pairings in 20 pigeons, but only in a condition where X-A training with new stimuli X was added; Velasco et al. (
2010), who reported a weak symmetry effect in one out of four pigeons when including partial symmetry training; and Lionello-DeNolf and Urcuioli (
2002), who found no symmetry in 12 pigeons, even when adding identity training.
Together, these six reports encouraged us to make a new attempt at demonstrating symmetry in non-humans with transfer effects. The aim of the present study, conducted in 20 Guinea baboons (Papio papio) and using a CMTS procedure, was to examine transfer effects on consistent versus inconsistent reversed pairings, while correcting for shortcomings of previous studies, notably: no additional training; both consistent and inconsistent conditions tested in all participants; control for the order of conditions; control of stimulus effect thanks to randomization across a large number of participants. We quantify here transfer effects on the reversed pairings in two ways: first test block performance (immediate transfer) and length of learning to criterion (global transfer). We provide responses on individual initial test trials, as a complementary measure.