Article Figures & Data
Figures
- Figure 1.
Rats with switching or non-switching experience had similar performance in a novel task. A, Experiment schematic for the differential experience phase. In this phase, all rats were trained for two sessions per day for up to 10 d. Non-switching group rats (H and 2T) learned a single task, either the H maze or the 2T maze, and they trained on the same maze during each training session, twice per day. Switching group (H/2T and 2T/H) rats learned both alternation tasks, H and 2T, counterbalanced for the order of maze sessions. Both the H and 2T mazes have four arm ends; two arm ends are reward locations (green circles). Visits to the other maze ends (white circles) are not rewarded. The rewarded visit sequence is shown. B, Experiment schematic for the common experience phase. All groups of rats learned to navigate the same plus maze task for two sessions per day. The task rule is an alternation sequence between Location 1 and Locations 2 or 3. The rewarded visit sequence is shown. C, The reward rate on the final session of differential experience phase. A two-way ANOVA did not show a significant effect on final performance from experience (p = 0.13) or task (p = 0.79) or their interaction (F(1, 1) = 0.70; p = 0.41). D, Proportion of trials rewarded per session on the plus maze grouped by experience and task in Phase 1. Pairwise Wilcoxon rank-sum test with Benjamini/Hochberg false discovery rate correction did not show a significant difference between experience (p > 0.44 for all pairs) or task (p > 0.83 for all pairs) across sessions.
- Figure 2.
Behavior choice pattern classification for the plus maze. A, Schematic of first-, second-, and third-order description of behavior choices, which were individual location visits, egocentric movements at junctions, and pairs of turns, respectively. B, Example behavior choice sequence and corresponding higher-order descriptions. Triangles indicate a transition between turn and straight actions. C, Example behavior choices for the first 30 trials in the plus maze for two animals in the non-switching group. First-order transitions shown by the circles that indicate the location visited by the rat. Red circles indicate the rewarded visits. Second-order transitions convert the location visit pairs into left turns (L), right turns (R), and straight (S). L and R are marked blue, and S is in orange. Triangles correspond to switch trials or third-order transitions that involve changes between L/R and S. Additional examples in Extended Data Figure 2-1. D, Example behavior choices for the first 30 trials in the plus maze for two animals in the switching group. E, Proportion of turn/straight transition trials (mean and 95% confidence interval of the mean) for each session. These are the frequencies of the trial pairs marked by triangles in C and D. Wilcoxon rank-sum test with Benjamin/Hochberg false discovery rate correction, p = 0.048 for the first session only. F, Proportion of transitions between left and right turns (mean and 95% confidence interval of the mean) for each session. Wilcoxon rank-sum test not significant. G, Proportion of repeated straight choices (mean and 95% confidence interval of the mean) for each session. Wilcoxon rank-sum test not significant. H, Proportion of repeated turn choices (mean and 95% confidence interval of the mean) for each session. Wilcoxon rank-sum test not significant.
- Figure 3.
Visualizing sequential behavior choice probabilities. A, Example dendrogram of conditional probabilities for three-trial choice sequences. Edges represent the conditional probability, and nodes represent the choices. B, Example showing the probabilities of sequences with one, two, or three trials. C, Choice probability dendrograms for the first 30 trials of the plus, H and 2T mazes. Three example animals from the non-switching (left column) and switching (right column) are shown. D, Scatter plot of the first three principal components of the choice probabilities between non-switching (black) and switching (red) groups for the first 30 trials on each maze. E, Cosine similarity across all principal components of choice probabilities between groups and within group for the first 30 trials. Wilcoxon rank-sum test p = 0.0022.
- Figure 4.
Models of non-switching and switching groups show distinct parameters that control how past trials influence future trials. A, Schematic of a modified ddCRP model. τ modulates the time-dependent influence of all previous trials on the next trial, effectively controlling the decay rate over all previous trials. Larger values of τ correspond to a more persistent influence of previous choices on the next choice. C modulates the dependence of the next trial on the immediate previous trial. Higher values of C correspond to a greater influence of the previous trial on the next trial. ɑ determines likelihood the next trial is drawn from a base distribution instead of trial history. β determines the likelihood the base distribution is governed by a uniform distribution (β = 0) or a distribution that is biased to repeat (β > 0) or avoid previous choices (β < 0). Model validation in Extended Data Figure 4-1. B, Scatter- and boxplots of model parameters from model fit to each animal's first 50 trials on the plus maze. Wilcoxon rank-sum test with Benjamini/Hochberg false discovery rate correction for τ (p = 0.023), C (p = 0.023), ɑ (p = 0.87), and β (p = 0.055).
- Figure 5.
Learned action patterns could not explain choice sequences in new task. A, Simulating choice sequences in Phase 2 based on learned action sequences from Phase 1. The previous two trials determined the distribution from which the next choice was selected. The second-order Markov models were trained on the observed action sequences from each maze in Phase 1. At each step, the model determined whether to select the next choice from the Phase 1 action distributions or from a uniform distribution (qE). If the Phase 1 distributions were chosen, the model selected the next choice from one of the two maze distributions (qs). B, Difference in probability of total turn/straight transition trials between simulation and observed behavior.
- Figure 6.
Non-switching and switching groups showed distinct errors during stable performance. A, Schematic of error types. Reference location errors were out-of-sequence choices after visiting Locations 2 and 3. Alternation errors were out-of-sequence choices after visiting Location 1. Errors were defined as the first out-of-sequence choice after correctly completing a sequence of four visits. B, Similar rates of all errors, both reference and alternation, for the last five behavior sessions. Wilcoxon rank-sum test p = 0.93. C, Non-switching group animals showed higher rates of alternation errors compared with switching group animals for the last five behavior sessions. Wilcoxon rank-sum test p = 0.027. D, Switching group animals showed higher rates of reference location errors compared with non-switching animals for the last five behavior sessions. Wilcoxon rank-sum test p = 0.00082.
- Figure 7.
Non-switching group errors consistent with a location-bias strategy. A, Expected error rates based on lapse trials following a location or action bias. The location bias assumes rats default to visiting the location most frequently rewarded in the task. This model predicted higher alternation compared with reference location error rates. The turn bias model assumed rats default to following a sequence of turns based on learned turn sequences from all tasks. These models predicted similar rates of alternation and reference location errors. Details of calculations in Extended Data Figures 7-1 and 7-2. B, Scatterplot of alternation versus reference location errors. C, Difference between alternation and reference location errors. The non-switching group showed a greater difference between alternation and reference location errors. Wilcoxon rank-sum test p = 0.0055.
Extended Data
Extended Data 1
Code for Chinese Restaurant Process model. Download Extended Data 1, ZIP file.
Figure 2-1
Behavior choices for the first 30 trials of the Plus maze for all animals, shown in the same format as Fig. 2 C-D. 1st order transitions shown by the circles that indicate the maze location visited by the rat. Red circles indicate the rewarded visits. 2nd order transitions convert the location visit pairs into left turns (L), right turns (R) and straight (S). L and R are marked blue, and S is in orange. Triangles correspond to switch trials, or 3rd order transitions that involve changes between L/R and S. Download Figure 2-1, TIF file.
Figure 4-1
Simulations confirm that the model parameters can be recovered from sequences of actions. We simulated from the distance-dependent Chinese restaurant process using two different sets of parameters (simulations 1 and 2, dashed lines indicate the true parameters). For each set of parameters, we generated 50 independent simulations. The parameters were then fit with an increasing number of trials using the posterior median as the estimate. The points give the mean estimates, and the error bars show a 90% interval over simulations. The estimated parameters remained close to the prior distribution with few trials and tended towards the true parameters with increasing amounts of data. We found that the context dependency parameter (C) required the fewest number of trials to separate across these two simulations. Given low values of the chosen base distribution bias (ɑ), which meant the base distribution was unlikely to be chosen in the generated sequences compared with the history-dependent distributions, we did not expect the repetition bias parameter (β) to be effectively recovered. Download Figure 4-1, TIF file.
Figure 7-1
Error likelihood predicted by a turn-bias strategy. Download Figure 7-1, TIF file.
Figure 7-2
Error likelihood according to a location-bias strategy. Download Figure 7-2, TIF file.
In this issue
