Higher-Order Conditioning: What Is Learnt and How it Is Expressed

Abstract

Pairing a neutral conditioned stimulus (CS) with a motivationally significant unconditioned stimulus (US) results in the CS coming to elicit conditioned responses (CRs). The widespread significance and translational value of Pavlovian conditioning are increased by the fact that pairing two neutral CSs (A and X) enables conditioning with X to affect behavior to A. There are two traditional informal accounts of such higher-order conditioning, which build on more formal associative analyses of Pavlovian conditioning. But, higher-order conditioning and Pavlovian conditioning have characteristics that are beyond these accounts: Notably, the two are influenced in different ways by the same experimental manipulations, and both generate conditioned responses that do not reflect the US per se. Here, we present a formal analysis that sought to address these characteristics.

Perceptual learning after rapidly alternating exposure to taste compounds: Assessment with different indices of generalization.

“…In fact, the impact of associatively activated representations on performance has been formally implemented in a recent application of a model of Pavlovian learning and performance (HeiDI; Honey et al, 2020) to higher-order conditioning (Honey & Dwyer, 2021, 2022. The supplementary finding reported by Recio et al (2018)-that placing a distractor between successive presentations of the compounds AX and BX (group Alternating) or AX and AX (and BX and BX; group Blocked) reduced the difference between the groups-might have reflected a disruption to the formation of a (direct) association between A and B in group Alternating, rather than to a process of perceptual learning (see cf.…”

Associative change in Pavlovian conditioning: A reappraisal.

“…These associations had already been implicated in phenomena that appeared to be beyond the scope of the Rescorla-Wagner model (e.g., potentiation : Durlach & Rescorla, 1980; unblocking : Rescorla & Colwill, 1983). Moreover, there is a significant body of independent evidence demonstrating their ubiquity, from studies of sensory preconditioning (e.g., Brogden, 1939; Rescorla, 1980; Rescorla & Cunningham, 1978) and second-order conditioning (e.g., Pavlov, 1927; Rescorla, 1982; Rizley & Rescorla, 1972; for a recent review, see Honey & Dwyer, in press, 2021). Indeed, recent evidence has revealed that such within-compound (or sensory) associations exhibit fundamental similarities to Pavlovian conditioning (e.g., Maes et al, 2020).…”

“…Equations 1 and 2 illustrate this rule in the context of Pavlovian conditioning trials, with the general form of the rule for any two stimuli (1 and 2) being: ΔV 1-2 = α 1 (c.α 2 − ∑V TOTAL-2 ). The consistent application of the rules generates associative chains (e.g., CS1→CS2→US), which can also contribute to performance (see Honey & Dwyer, in press, 2021).…”

“…This is the case because ∑V TOTAL-US will take a negative value (reflecting V CS2–US ), and the error term will be positive in: ΔV CS1-US = αβ(0−∑V TOTAL-US ). Of course, in the case of second-order conditioned inhibition (and excitation), the Rescorla-Wagner model needs to rely on the (informal) idea that associative chains (or indeed mediated learning) can outweigh the predicted effects on direct conditioning (for a formal application of HeiDI to higher-order conditioning; see Honey & Dwyer, in press, 2021). In the case under consideration here, the predicted excitatory association between A and B, for example, reflects the fact that on a reinforced A→US trial the US will inhibit B (see panel 2) and the error term for the change for the A→B association will be positive (i.e., 0–V US-B , with V US-B itself being negative).…”

“…6It is worth noting that the reduction in the difference between AD and BC, observed when the influence of the chains is included (panel 2), is not observed when the efficacy of the chains is modulated by the similarity of the perceived intensity of retrieved component (e.g., B via V A-B ) to its corresponding conditioned intensity (α B; Honey & Dwyer, in press, 2021). In the other simulations presented in this paper, however, such modulation reduces the size of the effect, with the result that the differences between AD and BC become more comparable when modulation by similarity is included throughout.…”