Rapid learning dynamics in individual honeybees during classical conditioning

Abstract

Associative learning in insects has been studied extensively by a multitude of classical conditioning protocols. However, so far little emphasis has been put on the dynamics of learning in individuals. The honeybee is a well-established animal model for learning and memory. We here studied associative learning as expressed in individual behavior based on a large collection of data on olfactory classical conditioning (25 datasets, 3298 animals). We show that the group-averaged learning curve and memory retention score confound three attributes of individual learning: the ability or inability to learn a given task, the generally fast acquisition of a conditioned response (CR) in learners, and the high stability of the CR during consecutive training and memory retention trials. We reassessed the prevailing view that more training results in better memory performance and found that 24 h memory retention can be indistinguishable after single-trial and multiple-trial conditioning in individuals. We explain how inter-individual differences in learning can be accommodated within the Rescorla-Wagner theory of associative learning. In both data-analysis and modeling we demonstrate how the conflict between population-level and single-animal perspectives on learning and memory can be disentangled.

Keywords: Apis mellifera; Rescorla–Wagner model; classical conditioning; learning curve; proboscis extension response (PER); single-trial learning; sucrose responsiveness; sucrose sensitivity.

Figure 1

Group-average CR probabilities do not adequately represent the CR probabilities in individual honeybees during classical conditioning of the proboscis extension response. (A) Binary conditioned response matrix from a typical dataset consisting of four conditioning trials and one memory retention test at 24 h (dataset 21). A gray entry indicates no CR, a black entry indicates a CR. Animals have been sorted according to their first CR and the number of consecutive CRs. (B) Average CR probability p(1) and conditional CR probabilities p(x_t = 1 | t_firstCR = 2) and p(x_t = 1 | t_firstCR = 3). Once animals have initiated their first response, they remain responding in subsequent trials with high probability. The dotted line indicates the time point in trial time t_max at which the regression curve (Equation 3) on the average CR probabilities assumes its maximum. (C) Histogram of first responses. The largest proportion of animals starts to respond on the second trial. (D) Binary conditioned responses matrix of a hypothetical dataset, which was generated by randomly permuting the CRs of dataset 21 across animals for each trial separately. (E,F) Analog analysis to (B,C). Group-average behavior represents individual behavior in the hypothetical dataset. Conditional probabilities do not reveal a serial dependency. The percentage of non-responders is drastically reduced.

Figure 2

Fast dynamics of associative learning during classical conditioning of the proboscis extension response. For each data set the upper panel shows the average CR probabilities and the CR probabilities in two subgroups of animals that start to respond on the second (t_firstCR = 2) or third trial (t_firstCR = 3). The black line depicts a regression curve (Equation 3) on the average CR probabilities (open square symbols). The dotted line depicts the position t_max of the maximum of the regression curve in trial time. The lower panel displays the percentage of animals that showed their first CR in a given trial. Animals that did not show a CR in any of the trials are represented by the black bar (none). Across all data sets, the largest proportion of animals starts to respond after only a single conditioning trial. Once animals have responded for the first time they have a high probability to continue responding in subsequent trials. The percentage of non-responding animals varies across datasets. Bees which responded to the first CS before the CS-US pairing were excluded from the analysis. (A) Dataset 4. (B) Dataset 6. (C) Dataset 21. (D) Dataset 10. (E) Dataset 11. (F) Dataset 12. (G) Dataset 13: The CR stability is decreased under trace conditioning (4.5 s gap between CS offset and US onset). (H) Data set 14: Control group for dataset 13 (CS and US overlap by 0.5 s). (I,J) Datasets 16 and 17: The dynamics of olfactory associative learning resemble the dynamics of tactile associative learning. (K) Dataset 18: The stability of the CR is decreased under massed training conditions (inter-trial-interval equals 30 s). (L) Dataset 19: The stability of the CR is high under spaced training conditions (inter-trial-interval equals 15 min).

Figure 3

Dynamics of discriminative learning during differential conditioning (data set 20). (A) CR probabilities to the CS+ and the CS− during the conditioning phase in three subgroups of animals, defined by their first response trial to the CS+ (t_firstCR = 1, 2, 3). Once animals have started to respond to the CS+, they have a high probability to continue responding to the CS+ in consecutive trials (curves with square markers). Animals responding to the CS+ (t_firstCR = 1, 2, 3) show high CR probabilities to the CS− in the first conditioning trials, and low CR probabilities to the CS− at the end of the conditioning phase (curves with round markers). CS− trials are indicted by an apostrophe. (B) CR probabilities to the CS+ and the CS− of the three subgroups at 1 and 24 h.

Figure 4

Effect of single-trial and multiple-trial conditioning on 24 h memory retention and discriminatory power under examination of individual CR histories during conditioning. ^*P < 0.05. (Ai) CR probability to the trained odor in subgroups 01, 0111, and 0000 of Experiment 1. Memory retention after four-trial and two-trial conditioning is not significantly different (0111 vs. 01 subgroup, χ² = 0.000960 with 1 degrees of freedom, P = 0.975). Animals that did never respond during four-trial conditioning showed poor memory retention. (Aii) Discrimination index (DI) in subgroups 01, 0111, and 0000. Discriminatory power of the memory after four-trial conditioning and two-trial conditioning is not significantly different (0111 vs. 01 subgroup, Mann–Whitney Rank Sum Test, T = 10379.000, P = 0.095). Animals that did not respond during four-trial conditioning show poor memory discrimination. (Aiii) Duration of the proboscis extension to the trained odor in subgroups 01, 0111, and 0000. The CR duration is not significantly different after four-trial and two-trial conditioning (0111 vs. 01 subgroup, Mann–Whitney Rank Sum Test, T = 11901.500, P = 0.238). (Aiv) Discrimination Index computed on CR duration (DI_dur) in subgroups 01, 0111, and 0000. The CR duration does not reveal significant differences in memory discrimination after four-trial and two-trial conditioning (0111 vs. 01 subgroup, Mann–Whitney Rank Sum Test, T = 10652.500, P = 0.248). (Bi) Memory retention after two-trial and single-trial conditioning is not significantly different (Experiment 2, 01 vs. 0(1) subgroup, χ² = 2.935 with 1 degrees of freedom, P = 0.087). (Bii) The discrimination index after two-trial and single-trial conditioning is not significantly different (01 vs. 0(1) subgroup, Mann–Whitney Rank Sum Test, T = 8346, P = 0.146). (Biii) The duration of the proboscis extension response to the trained odor is significantly longer after two-trial than after single-trial conditioning (01 vs. 0(1) subgroup, Mann–Whitney Rank Sum Test, T = 7265.5, P ≤ 0.001). (Biv) The duration of the proboscis extension response did not reveal significant differences in discriminatory power after two-trial and single-trial conditioning (01 vs. 0(1) subgroup, Mann–Whitney Rank Sum Test, T = 8256.5, P = 0.093). (Ci) Animals that started to respond early during conditioning showed significantly more memory retention than animals that started to respond later during conditioning [Experiment 1, 01 vs. (001, 0001) subgroup, χ² = 7.246 with 1 degrees of freedom, P = 0.007]. (Cii) Early and late responders do not significantly differ in memory discrimination (Mann–Whitney Rank Sum Test, T = 5489, P = 0.346). (Ciii) The duration of the proboscis extension response to the trained odor is significantly longer in early than in late responders (Mann–Whitney Rank Sum Test, T = 4925.5, P = 0.016). (Civ) The duration of the proboscis extension response does not reveal significant differences in memory discrimination between early and late responders (Mann–Whitney Rank Sum Test, T = 5428, P = 0.271).

Figure 5

Injection of the translation blocker emetine before conditioning had no effect on 24 h memory retention. Bees received a saline or emetine injection 30 min before single- or three-trial conditioning. Memory retention was tested 24 h after conditioning. ^***P < 0.001. (A,B) Averaged performance of all bees; (C,D) averaged performance of bees which showed a conditioned response (CR) during training. (A) The percentage of bees which showed a CR in the 24 h memory retention test differed between single-trial and three-trial conditioning [F_{(1, 289)} = 45, p < 0.001; Two-Way ANOVA] but not between emetine- and saline-injected bees [F_{(1, 289)} = 0.9, p = 0.34]. (B) The discrimination index (DI) of the 24 h memory retention test neither differed between single- and three-trial conditioning [F_{(1, 289)} = 3.3, p = 0.07] nor between emetine- and saline-injected bees [F_{(1, 289)} = 0.2, p = 0.66]. (C) The percentage of bees which showed a CR in the 24 h memory retention test differed between single- and three-trial conditioning [F_{(1, 191)} = 21.2, p < 0.001] but not between emetine- and saline-injected bees [F_{(1, 191)} = 0.1, p = 0.8]. (D) The DI of the 24 h memory retention test neither differed between single- and three-trial conditioning [F_{(1, 191)} = 0.2, p = 0.7] nor between emetine- and saline-injected bees [F_{(1, 191)} = 0.2, p = 0.6].

Figure 6

Sucrose responsiveness correlates with learning performance in olfactory and tactile classical conditioning (compare to Figures 1, 2 in Scheiner et al., 2001a). (A) Average CR probabilities in four subgroups of animals from olfactory conditioning (dataset 16). Animals were divided into subgroups on the basis of individual gustatory response scores (GRS). Small numbers indicate low responsiveness to sucrose. Animals with high responsiveness to sucrose reach higher plateaus in CR probability than animals with low responsiveness to sucrose. (B) Histogram showing the percentage of animals for each subgroup that start to respond in a given trial. Animals showing at least one CR in any of the trials start to respond early during conditioning. Most non-responders have a low responsiveness to sucrose. (C) Average CR probabilities in four subgroups of animals from tactile conditioning (dataset 17). (D) Histogram showing the percentage of animals for each subgroup that starts to respond in a given trial. The dynamics of tactile learning resemble the dynamics of olfactory learning.

Figure 7

The extended version of the Rescorla–Wagner model RW^{P(α, λ)}. (A) The gradual increase of associative strength v_t across training trials in an individual animal is described by two parameters: the rate of learning α and the asymptotic performance level λ. The probability for emitting a CR on a given trial is assumed to equal the associative strength. (B) Color-coded joint distribution P(α, λ) of learning parameters of all animals from dataset 21. Most animals have high learning rates and high learning asymptotes. The inlets on the top and on the right side of the colored joint distribution depict the cumulative probability distributions P(α) and P(λ) for which the joint distribution P(α, λ) has been summed over lambda or alpha, respectively. (C–F) Color-coded likelihoods for observing a particular sequence of CRs given different combinations of α and λ. The leftmost number of the CR sequence equals the CR in the first trial, and the rightmost number in the CR sequence equals the CR in the last trial.