Article Figures & Data
Figures
- Figure 1.
The Hex Maze. A, Shows the maze with intramaze and extramaze cues (left) and the maze from the view of the mouse (right; also see Movie 1). B, The main performance metric is the log-normalized path (pathnorm), with the lengths of the paths taken by the animal divided by the shortest possible path to the GL (indicated by the X). Thus, for all subsequent figures the number in brackets of the log is the relative length of the path taken by the animal, with 2 indicating that the path was twice as long as the shortest possible path. C, During training, animals started each trial from a different location and had to navigate to a fixed GL. A first trial measures long-term memory performance and was used as a probe trial on critical sessions (no food present). Performance on all trials of the session measure general working memory/navigational performance in the known environment. D, After the animals had acquired the general maze knowledge during the Build-Up, Updates were performed with inclusion of new Bars, new goal Locs, or the inclusion of L + B. E, The general training schedule for all animals during the whole experiment. Animals were trained to one GL in a given session. For group 1, the GL was kept constant for seven sessions of GL1, then five or six sessions for GL2, and five or seven sessions each for GL3–5. Additionally, three of the initial five locations were repeated with each of three sessions. For group 2, the GL was kept constant for seven sessions of GL1, then five sessions for GL2–3, and seven sessions each for GL4. Finally, for all cohorts, each Update contained three sessions. The sequence of the Update types was counterbalanced across animals (session 1 of each update indicated with an arrow). Each Update type was repeated two to three times. Throughout all phases, the first trial of the second session and during Build-Up first trial of the fourth, fifth, or sixth session were used as probe trials. Group 1 was trained 3 d/week (3dw), group 2 was trained 2 d/week (2dw). F, G, Example paths of the Build-Up (F) and Updates (G) are shown (Movies 2, 3, 4, 5, 6, 7, 8, 9). T1, Trial 1. Data are in Extended Data Figure 1-1.
- Figure 2.
HexMaze performance group 1 3 d/week training. A, Shows schedule examples for the Build-Ups and Updates. Orange boxes indicate days with probe trials (no food for 60 s of the first trial). B, Performance across all trials (including first trial) measures general working memory/navigational performance within the environment. During Build-Up, there was a significant effect across session and across the five GL switches. In contrast, during Updates, only if a location switch was involved in the update (Loc/L + B), performance was worse during the first session of the change and an improvement across sessions is visible. C, Performance on the first trial of each session measures the ability to remember the GL from 2–3 d ago. During the Build-Up, long-term memory improved across sessions. During the Updates, there was an improvement across sessions as well as a difference between types with larger changes in the environment (linear from Bar to both L + B), leading to worse performance. This is especially noticeable in session 1 for Loc and L + B switches where the goal is initially unknown, whereas for a Bar update only an adaptation of the route is involved. *p < 0.05, **p < 0.01. Error bars are the SEM. The number in brackets of the log is the relative length of the path taken by the animal (taken path T/shortest path S), with 2 indicating that the path was twice as long as the shortest possible path.
- Figure 3.
HexMaze group 2 performance 2 d/week training. A, Shows schedule examples for the Build-Up. The schedule for group 2 is shown in purple, and for group 1 in light green, illustrating the resulting shift in alignment of training days and time. Orange boxes indicate days with probe trials (food not present for 60 s of the first trial). B, Performance across both the all-trials measurement (general working memory/navigational performance within the environment) and the first-trial measurement (long-term memory). We found a significant improvement in performance across sessions for both measures and additionally across GL and GL × session interaction for all trials. C, To compare 2 with 3 d/week training, we included the corresponding training day as well as session according to the time of group 1 and compared these with the performance of group 2. It is important to note, that the performance depended on how much time had elapsed since first exposure to the maze (weeks), not how much training the animals had received (TD, training day). D, Examples from the study schedule of the Updates. With 2 d/week a natural alternation of 2 and 5 d, gaps ensued during the Updates. E, Comparing only the 2 d Updates of group 2 with the Updates of group 1 (also 2 d gaps) showed only an Update difference during the first session. F, Plotted is the performance during Updates for group 2 for both the 2 and 5 d delays. One session of training only led to significant long-term memory that lasted 2 d not 5 d, whereas two training sessions did indeed lead to a 5 d memory persistence visible in the third session (2 d condition for session 2). *p < 0.05, **p < 0.01, ***p < 0.001. Error bars are the SEM. The number in brackets of the log is the relative length of the path taken [taken path (T)/shortest path (S)] by the animal, with 2 indicating that the path was twice as long as the shortest possible path.
- Figure 4.
Three phases of map learning. A, Plotted are all trials separated for the four GLs during Build-Up as well as the different Update types with separate lines for first session, second session and third session onwards (for Build-Up, it is S3–5/7; for Updates, it is just S3 since no further sessions were run). Three learning phases are noticeable: learning the first goal location, learning the second goal location with better session 2 performance, and learning the Updates with already good session 1 performance. B–E, The first trial of session 1 (B), the last trials of session 1 (C), the first trials of session 2 (D), and the change from the last trial of sessions 1 to the first trials of session 2 (E) are shown for the different goal locations during Build-Up (GL1–4 as well as the different update types). The first trial performance during session 1, when the goal is unknown, first became worse in GL2–4 compared with GL1, most likely because of animals first navigating to the old goal location. Only in the Barrier updates (light blue) was performance better than in all other GLs and updates since the location did not change. At the end of session 1 (last trial), there is no difference between the different GLs and updates. The three phases of learning are again noticeable in the first trial of session 2, reflecting long-term memory after one session of training. This showed a stepwise function, improving in GL2–4 in contrast to GL1 and improving even more during the updates. The same is reflected in the difference values presented in E (updates; one-sample t test to 0: t(71) = 4.2, p < 0.001).
- Figure 5.
Previous knowledge effects. In these panels, we highlight some previous knowledge effects A, The whole session performance for the first GL during the Build-Up. The significant session effect reveals a performance increase dependent on experience, indicating a more efficient working memory/navigational performance. B, Plots the performance for the first two sessions of the first two GLs during the Build-Up, as well as Updates (averaged for all types). Already for the second GL (3 weeks since training start), a significant increase in performance (decrease of path length) is seen in the second session compared with the first session. This overnight (offline) performance increase is comparable to the increase found after seven sessions for the first GL. This may represent a more efficient consolidation and updating effect, but is expressed only in the whole session average (not long-term memory present in the first trial; Figs. 2, 3). During the Updates, this performance increase is already visible in the first session with additional offline gains found in the second session. This three-step performance gain is reminiscent of a learning-set effect (Harlow, 1949). C, Considering all sessions, we find that animals already reach overall plateau performance by the second GL. D, Zooming in on the performance during the first and second session during the Updates, another previous knowledge effect is revealed across the different Update types. The Bar, Loc, and L + B differed in their overlap of previous knowledge (or need for updating that knowledge), which influenced how well they performed (all-trial) in the first session. E, The same effect but now only for the first trials. Only in the presence of a goal switch did performance in the first session decrease. However, by the second session this performance difference was gone, revealing that one session is sufficient for the memory update. F, The performance of only the first trial of the second session during the Build-Up and Updates (only Loc and L + B) where long-term memory (2–3 d) after one session of learning a new GL improves from Build-Up to Updates. Thus, it seems that once a cognitive map is established, only one session of training leads to better long-term memory performance. Orange boxes indicate that the trial was used as a probe trial, meaning food was not present for the initial 60 s. *p < 0.05, **p < 0.01, ***p < 0.001. Error bars are the SEM. Data were taken from both groups 1 and 2. The number in brackets of the log is the relative length of the path taken by the animal [taken path (T)/shortest path (S)], with 2 indicating that the path was twice as long as the shortest possible path.
- Figure 6.
Probe trial analysis (each session 2, trial 1). Across the goal location switches during Build-Up and during the Updates, an increase in the number of crossings could be seen for both the current and previous goal locations compared with the two control nodes (groups 1 and 2, GL2–4 and Loc for full model, n = 16; node: F(3,45) = 22.3, p < 0.001; GL2–4/Loc: F(3,45) = 10.7, p < 0.001; interaction: F(9,135) = 2.0, p = 0.044).
- Figure 7.
Changes because of updates. A, B, Always the last trial before an update (S3 of previous condition) and the first trial of the update (A) and the difference values (subtraction) for these (B). From left to right the shortest possible path, the taken path, and the relative path are presented. If barriers were included (Bar and L + B), the shortest possible path would increase from the previous trial. But only if location was changed (Loc and L + B) did the taken path increase, for the Bar update the taken path only increased by the same amount of the shortest path (two nodes). Interestingly, because of the change in shortest path, the relative change (taken/shortest) actually decreased in the Bar update. #p = 0.087, **p < 0.01, ***p < 0.001; A, paired t tests; B, one-sample t test to 0. Data taken from both groups 1 and 2. Error bars are the SEM. The number in brackets of the log is the relative length of the path taken by the animal [taken path (T)/shortest path (S)], with 2 indicating that the path was twice as long as the shortest possible path.
- Figure 8.
Within-session learning. A–E, The change in performance within session (trial 1 and then blocks of 10 trials) always across the first three sessions (S1–3), the very first goal location (A), averaged across the subsequent goal locations of the buildup (B), for barrier updates (C), for location updates (D), and for location and barrier (Loc + Bar) updates (E). The first goal location did not show strong within-session learning during these first sessions; in contrast, later on (B) the main learning occurred between trial 1 and the next trial block during session 1 and trial 1 of sessions 2 and 3 started lower, but additional within-session improvement could be seen in the next block. In the barrier updates, performance was starting trial 1 of first session well and remained stable. For the other updates, a linear improvement during session 1 was seen across trials, and now performance was sustained to sessions 2 and 3 with no strong additional gains from the first trial to subsequent trials. For statistics, see the main text. Data were taken from both groups 1 and 2. Error bars are the SEM. The number in brackets of the log is the relative length of the path taken by the animal [taken path (T)/shortest path (S)], with 2 indicating that the path was twice as long as the shortest possible path.
- Figure 9.
Different performance parameter buildup and updates. Shown are the log of normalized path length (top row, as used throughout the article), normalized path length, percentage of trials that was a direct run (second from top), percentage of trials that was a direct run after the second node since mice often would initially run in heading direction and then stop to consider where to go (third from top). As a final parameter, we took for each node whether the choice would bring the mouse closer to the food (correct) or not (incorrect) and created an average per trial across all traversed nodes. Each left, all trials; right, first trials for Build-Up (Group 2) and Updates (all). Error bars are the SEM. Lines are polynomial fits. The same effects seen in the log of the normalized path length can also be seen in the other parameters.
Movies
- Movie 1.
Mimicking animal view in the maze.
- Movie 2.
Three trials of S1 during the Build-Up Mouse 12.
- Movie 3.
Three trials of S2 during the Build-Up Mouse 12.
- Movie 4.
Three trials of S1 during the Build-Up Mouse 17.
- Movie 5.
Three trials of S2 during the Build-Up Mouse 17.
- Movie 6.
Three trials of S1 during the Update Mouse 6.
- Movie 7.
Three trials of S2 during the Update Mouse 6.
- Movie 8.
Three trials of S1 during the Update Mouse 19.
- Movie 9.
Three trials of S2 during the Update Mouse 19.
Figure 1-1
Data from all trials used in all figures. Download Figure 1-1, XLSX file.
In this issue
