Embryo learning phase
The embryo learning phase took between 29 and 49 sessions to meet criteria and move on to the next phase (for experimental details see supplementary table 1). All eight pigeons performed two consecutive sessions at 80% correct of the free-choice trials before moving to the next phase. After reaching the initial learning criterion, a step-wise reduction of forced-choice trials as well as of reward probability followed to move to the transfer test phase.
Transfer test phase
All eight pigeons were successfully tested on known- and transfer-stimuli during the test session when transfer trials were rewarded non-differentially. Overall, performance in known-stimuli and transfer-stimuli trials was very high (Mean 91.81%, SD 3.26%; Mean 91.36%, SD 3.80%, respectively, see Fig. 4) and was significantly different from chance [t(7) = 4260.45, p < 0.001, Cohen’s d = 12.83; t(7) = 3.654.37, p < 0.001, Cohen’s d = 10.89, respectively]. There was no difference in performance between known- and transfer-stimuli trials [t(7) = 0.61, p = > 0.250, Cohen’s d = 0.45]. Individually, the two stimulus classes, both class X (known-stimuli trials mean 92.89%, SD 3.23%; transfer-stimuli trials mean 92.58%, SD 3.03%) and class Y (known-stimuli trials mean 91.26%, SD 3.98%; transfer-stimuli trials mean 90.13%, SD 6.55%), showed above chance-level performance (all ps < 0.001). There was no difference between stimuli of class X and Y in both known- [t(7) = 0.97, p > 0.250, Cohen’s d = 0.32] and transfer-stimuli trials [t(7) = 1.02, p = > 0.250, Cohen’s d = 0.42]. Behavioral results of the four pigeons who completed the transfer test in the absence of reward are presented in the supplementary results.
We used a k-nearest neighbor classifier to identify whether the presented stimulus class can be predicted based on the pecking location. To this end, the classifier was trained with randomly chosen pecking events across behavioral sessions in which pecking events were labeled according to the presented stimulus class in a given trial. Another subset of trials was then used to determine the predictive accuracy of the classifier based on pecking location alone (see “Methods”). This was done for each animal separately. Results of all analyses for the four pigeons that successfully performed the non-reinforced transfer are presented in the supplementary data.
Our aim was to investigate whether the class label can be predicted in correct–correct or CC classification in which both training and test trials were constituted by correct trials. All eight animals could be investigated for a CC classification in known-stimuli and transfer-stimuli trials. For known-stimuli trials, we found that the classifier could significantly predict the stimulus class based on pecking location for seven out of eight pigeons (P578: 82.08%, P580: 80.84%, P582: 70.56%, P583: 64.24%, P592: 77.84%, P593: 64.64%, P599: 77.80%, all ps < 0.001; for the classification accuracy based on shuffled input data see supplementary Fig. 11). Only results for pigeon P579 did not differ significantly from chance [52.40%, t(9) = 1.22, p > 0.250, Cohen’s d = 0.29, see Fig. 5] as the animal indifferently pecked on one side of the stimulus irrespective of the presented stimulus class.
Animals preferred different embryo features indicated by their idiosyncratic pecking locations for the two stimulus classes (Fig. 6).
Since individual pecks often came from the same trial, we validated our analysis only using single pecks from a particular trial to rule out that our results were influenced by possible dependencies of pecks within a trial. To this end, we only chose one random peck event from each trial as a valid event from which we randomly drew 250 training and 250 test pecks for classification. Results were virtually indistinguishable from our previous approach using all pecking events indicating that this approach was valid and not influenced by peck dependencies within the trial (P578: 81.28%, P579: 51.88%, P580: 80.92%, P582: 69.96%, P583: 63.84%, P592: 76.48%, P593: 67.12%, P599: 79.68%, see supplementary Fig. 12). For that reason, all following analyses were computed including all pecks from all trials to increase the number of available data points.
In a secondary analysis, we were also interested whether the first peck was already sufficient for stimulus classification as has been demonstrated by Cook et al. (2005) in a spatial choice task. Interestingly, using only the first peck for classification resulted in overall worse classification accuracy compared to the following pecks (see Fig. 7). For almost all pigeons, classification accuracy increased with consecutive pecks on the sample stimulus indicating that features of interest were not immediately pecked on after stimulus onset.
For transfer-stimuli trials (Fig. 5), we found identical results compared to known-stimuli trials (P578: 82.09%, P579 = 53.48%, P580: 83.68%, P582: 65.52%, P583: 64.96%, P592: 76.12%, P593: 63.16%, P599: 76.20%, all ps < 0.001). There was no difference in classification accuracy from known-stimuli trials in CC accuracy for any pigeon.
For all eight animals, there were sufficient errors allowing for a correct–error or CE classification in known-stimuli trials in which the classifier was trained with correct trials and tested on error trials. Here, two pigeons demonstrated similar pecking patterns in correct and error trials as the classifier yielded above chance results in a CE classification (P582: 62.36%, P599: 61.16%, all ps < 0.001, Fig. 8A, for the classification accuracy based on shuffled input data see supplementary Fig. 13). Pecking in error trials was more dispersed compared to correct trials resulting in a reduced accuracy for CE compared to CC classification for both pigeons (all ps < 0.001). This finding was, however, not exclusive for animals that pecked on similar locations during correct and error trials. Seven out of eight pigeons demonstrated significant differences when comparing the cumulative distributions of pecking (pooled across class X and Y) across the stimulus squares between correct and error trials (578: D = 0.12, p = 0.003, 579: D = 0.19, p < 0.001, 580: D = 0.47, p < 0.001, 582: D = 0.35, p < 0.001, 583: D = 0.35, p < 0.001, 592: D = 0.19, p < 0.001, 599: D = 0.49, p < 0.001, see supplementary Fig. 14). Results for class X and Y individually can be found in supplementary Table 2.
For pigeons P578 and P592, we found that CE classification was significantly below chance (P578: 38.12%, P592: 38.96%, both ps < 0.001). This type of pecking behavior indicates that the pigeons showed the “wrong” behavior for the respective stimulus class (e.g. they pecked on class Y features during class X trials and vice versa). Thus, they recognized the stimulus classes, but mixed up their identity in error trials.
The last three pigeons demonstrated no significant difference from chance in the CE classification (P579: 49.16%, P583: 49.00%, P593: 52.24%, all ps > 0.174) indicating that there was no relationship between correct and error trial pecking locations. For two pigeons, sufficient errors were made in transfer-stimuli trials enabling further analysis. For P578, we again found a significantly lower classification accuracy compared to chance [29.64%, t(9) = 12.64, p < 0.001, Cohen’s d = 3.97, Fig. 8B]. For pigeon P593, there was no association between correct and error trial pecking locations [52.70%, t(9) = 1.55, p = 0.156, Cohen’s d = 0.49]. Both pigeons, thus, showed consistent patterns in known-stimuli and transfer-stimuli trials.
Error–error or EE classification in which the classifier was both trained and tested on pecking events during erroneous trials could be performed for all animals in known-stimuli trials. Aim of the analysis was to quantify the internal consistency in pecking during error trials. It was above chance for five of the eight animals (P580: 63.16%, P582: 61.64%, P583: 74.44%, P592: 57.52%, P599: 72.96%, all ps < 0.004) indicating that pecking locations in error trials were to a certain degree consistent rather than random. A consistent response pattern in error trials was predominantly found in birds in which the C–E classification also yielded above or below chance level results, validating our interpretation of the pecking behavior during error trials. For pigeons P578, 579 and P593, no significant difference from chance could be found (all ps > 0.025). For pigeons P578 and P593, there were sufficient error events in transfer-stimuli trials to allow for EE classification. Interestingly, classification accuracy was above chance for pigeon P578 [67.6%, t(9) = 5.95, p < 0.001] in known-stimuli trials. For pigeon 593, it remained at chance level [53.92%, t(9) = 1.29, p = 0.228].