Flexible use of memory by food-caching birds

Abstract
Editor's evaluation
eLife digest
Introduction
Results
Discussion
Materials and methods
Data availability
References
Article and author information
Metrics

Abstract

Animals use memory-guided and memory-independent strategies to make navigational decisions. Disentangling the contribution of these strategies to navigation is critical for understanding how memory influences behavioral output. To address this issue, we studied spatial behaviors of the chickadee, a food-caching bird. Chickadees hide food in concealed, scattered locations and retrieve their caches later in time. We designed an apparatus that allows birds to cache and retrieve food at many sites while navigating in a laboratory arena. This apparatus enabled automated tracking of behavioral variables – including caches, retrievals, and investigations of different sites. We built probabilistic models to fit these behavioral data using a combination of mnemonic and non-mnemonic factors. We found that chickadees use some navigational strategies that are independent of cache memories, including opportunistic foraging and spatial biases. They combine these strategies with spatially precise memories of which sites contain caches and which sites they have previously checked. A single memory of site contents is used in a context-dependent manner: during caching chickadees avoid sites that contain food, while during retrieval they instead preferentially access occupied sites. Our approach is a powerful way to investigate navigational decisions in a natural behavior, including flexible contributions of memory to these decisions.

Editor's evaluation

The extreme memory capacities of food-caching birds provide untapped opportunities for studying mechanisms of memory formation and retrieval. Here, Applegate and Aronov develop an automated animal and cache-site tracking system in which moments of seed deposits, retrievals, and checks are measured continuously alongside the animal's spatial positions. Probabilistic models reveal idiosyncratic spatial preferences in individual birds and also identify flexible memory usage – in which a single memory of past seed deposition can differentially guide spatial trajectories depending on if the bird is in engaged in retrieving or storing seeds. The rigorous behavioral tracking and modeling sets the stage for dissection of neural mechanisms underlying memory storage and retrieval.

https://doi.org/10.7554/eLife.70600.sa0

eLife digest

Humans form new memories about what is happening in their lives every day. These autobiographical memories depend on a part of the brain called the hippocampus. But how these memories are recorded remains unclear. Studying certain birds may help to provide more insight.

Black-capped chickadees, for example, are memory specialists. They stash thousands of food items and use their memories to recover these hidden food stores. This behavior also relies on these birds’ hippocampus. Studying these animals' behavior in the laboratory may help scientists decode how the birds use their memories and to gain more insight about the brain processes underlying memory.

Now, Applegate and Aronov show that chickadees use memory not only to retrieve food but also to decide where to hide it in the first place. In the experiments, chickadees were placed in a specialized enclosure with a grid of holes covered by silicone rubber flaps on the floor. The birds lifted the flaps with their toes or beak to hide a piece of sunflower seed underneath. Applegate and Aronov recorded and analyzed the animals’ seed hiding and retrieving behavior with a video camera to determine whether the birds were remembering the sites or happening on them by chance.

This revealed that black-capped chickadees use the same memories of where they had hidden food in two different ways. When they were hiding new morsels, the birds remembered where they had stashed food and avoided those flaps. When they were retrieving food, the birds knew exactly which flaps to look under. Future experiments using this special enclosure may help scientists monitor what happens in the chickadees’ brains during these activities.

Introduction

The study of episodic memory in animals has been greatly benefitted by experiments that involve spatial navigation (DeVito and Eichenbaum, 2010; Morris et al., 1982; Salwiczek et al., 2010). Whereas memory use is internal to the animal and not directly observable, navigational decisions that require memory are overt and trackable. In addition, memory-related neural circuits, like the hippocampus, exhibit well-described firing patterns that correlate to navigational variables – including place and head direction (O’Keefe and Dostrovsky, 1971; Taube et al., 1990). However, navigation is a rich behavior that includes both mnemonic and non-mnemonic spatial strategies. For example, an animal using memory to obtain a reward may also engage in opportunistic foraging, use memory-independent search strategies, and exhibit spatial biases (Dally et al., 2006; Kamil and Roitblat, 1985; Krebs et al., 1974; Lamprea et al., 2008; Pravosudov, 2001). It is critical to tease apart these contributions to behavior in order to eventually understand the underlying neural mechanisms.

Food-caching birds offer an opportunity to study memory in the context of navigation. These birds hide food in many distinct, concealed locations throughout their environment and later retrieve their caches. Food caching has been studied in two general types of settings well-suited for characterizing different components of the behavior. In some experiments, spatial navigation is minimized, and birds choose from a small number of available options to obtain food (Brodbeck et al., 1992; Clayton and Dickinson, 1998; Clayton and Krebs, 1994a; Salwiczek et al., 2010). These tasks have been ideal for characterizing various contributions to memory, including location, content, and even the relative time of different caches. In contrast, other experiments have been performed in the wild or in large, naturalistic settings (Cowie et al., 1981; Herz et al., 1994; Krushinskaya, 1966; Sherry, 1984a; Stevens and Krebs, 2008)^. These studies have been well-suited for measuring spatial properties of the food-caching behavior, including the distribution of caches, large-scale biases, and the spatial specificity of cache memories. However, it is unknown how mnemonic and other spatial strategies are coordinated by the animal within a single behavior.

Addressing this question may be possible in reduced spatial settings, like those typical in rodent experiments (Muller et al., 1987; Poucet et al., 2003; Tolman, 1948). Such a setting must retain key spatial aspects of bird navigation, while also incorporating food-caching behavior and memory use. Ideally, it would also permit detailed tracking of the animal’s behavior and be compatible with neural recordings. We developed an experimental setup to fulfill these requirements for a food-caching species, the black-capped chickadee. We then used this setup to dissect the contributions of mnemonic and non-mnemonic navigational strategies to food caching, using probabilistic modeling of behavioral choices (Brunton et al., 2013; Raposo et al., 2012; Scott et al., 2017).

Results

Design of the behavioral paradigm

Our first goal was to engineer a behavioral setup in which chickadees navigate, cache food, and retrieve caches. We designed individual cache sites as holes in the floor of an arena covered by silicone rubber flaps (Figure 1). Dimensions and materials were chosen to allow chickadees to pull the flap open with little effort – using either their toes or their beak – and to deposit a piece of a sunflower seed underneath. Once released, the flap obstructed any visual access to the contents of the site from above. Flaps ensured that no visually guided, memory-independent strategy could be used by the chickadee to determine site contents. Sixty-four cache sites were arranged in an 8 × 8 grid inside a 61 × 61 cm square arena. In addition, the arena contained a food source in each of the quadrants: a feeder with a motorized cover that could be individually opened or closed. This setup allowed navigation on a similar spatial scale to that typically studied in rodents (Poucet et al., 2003) while also offering birds a large number of concealed cache sites.

Figure 1 with 2 supplements see all

Download asset Open asset

Behavioral paradigm for food caching and retrieval in chickadees.

(A) Rendering of the behavioral setup. For clarity, the front wall that contains doors of the arena is removed. Pink box highlights one cache site.(B) Illustration of a chickadee at a cache site. Top: chickadee prior to caching. Middle: cross-section of the cache site, with a chickadee pulling open the silicone flap to deposit a seed. Bottom: cross-section of the cache site after the seed was cached. (C) Left: video frames from the bottom camera showing all 64 cache sites at four timepoints within a behavioral session. Right: real-time detection of cached seeds. Red area indicates shape enclosed by the detected seed contour. (D) Detection of events at one example cache site in one behavioral session. Top: cartoon of the bird’s interactions with the site at several time points. Middle: video frames from the bottom camera at the corresponding time points. Bottom: Pearson correlation of each video frame with the image of the same cache site when empty. Caches create sustained decreases in the correlation, whereas site checks create transient decreases. (E) Ethogram of all behavioral events in an example session of the Caching task. Colored circles correspond to caches, retrievals, and site checks, as in (D). (F) Same as (E), for the Retrieval task. Colored regions indicate the phase of the trial. Black vertical lines denote trial boundaries.

We next developed methods to automatically track chickadee behavior in the caching arena. Automated behavioral tracking was necessary to allow closed-loop manipulations for some of the experiments described below. To achieve this, we used a transparent material for the bottom layer of the arena and positioned a video camera underneath. We applied a real-time contour detection algorithm (Lopes et al., 2015, see Materials and methods) to determine whether each of the sites was empty or occupied by a seed (Figure 1C). During offline analysis, we also detected instances of the bird opening the cover flap of a site, which we call ‘checks’. To do this, we calculated the Pearson correlation between the image of each site when empty and every video frame of that site in the session. Periods of time when a site was occupied corresponded to sustained periods of low correlation, while checks corresponded to transient decreases in correlation (Figure 1D). Thus, caches, retrievals, and checks could all be detected in the behavior using a single-camera video recording. We also trained a deep neural network (Mathis et al., 2018; Nath et al., 2019) to track the bird’s location for this analysis.

We designed two tasks, each suited for analyzing different aspects of behavior (for details see Materials and methods). In the Caching task (Figure 1E, Figure 1—video 1), chickadees were free to cache without any imposed trial structure for 1 hr. Individual feeders opened and closed to motivate movement through the arena and caching, but the schedule of these openings and closings was unrelated to the bird’s behavior. Our term ‘Caching task’ does not mean that birds only cached in this task – in fact, birds also ate seeds, checked sites, and retrieved their caches. However, because this task placed no limit on the number of caches, it was particularly well-suited for investigating the chickadee’s choice of where to cache each seed. This task was performed by 17 chickadees. In 10 of these, we recorded enough caches (> 64; see Materials and methods) for inclusion in the remainder of the Results section.

In the Retrieval task (Figure 1F, Figure 1—video 2), a trial structure was imposed. At the beginning of each trial, one of the feeders was open (‘feeder-open phase’), allowing chickadees to eat or cache seeds. Once the bird cached 1–3 seeds, lights were turned off for a 2 min delay phase. After the delay, all feeders were closed (‘feeder-closed phase’), and the only sources of food in the arena were previously cached seeds. In the feeder-closed phase, birds were generally motivated to find and retrieve their caches. This task was therefore suitable for investigating which sites the chickadee choses to check during cache retrieval. This task was performed by seven chickadees, all of which are included in the analyses.

Chickadees have unique and stable spatial biases

We first asked whether chickadees cached into some sites of the arena more than into others. In the Caching task, we calculated the cache distribution ( $p^{b i a s}$ ) by dividing the number of caches into each of the 64 sites across sessions by the total number of caches (Figure 2A). Birds used most of the arena for caching (between 41–64 sites for each of the 10 birds), but in many cases cache distributions appeared non-uniform. To quantify this, we measured the entropy of $p^{b i a s}$ for each bird and compared its value to simulations in which the same number of caches were sampled from a uniform distribution (Figure 2B). In all birds, entropy was lower than expected by chance (p < 0.001). Therefore, birds exhibited spatial biases, even though they cached widely throughout the arena.

Figure 2 with 2 supplements see all

Download asset Open asset

Chickadees exhibit idiosyncratic and stable biases in the locations of caches.

(A) Probability distributions of cache locations across all sessions for four example birds. Distributions are denoted by $p^{b i a s}$ in the text. (B) Red lines: entropy values of the spatial distributions shown in (A). Gray histograms: entropy values for simulated caches drawn from a uniform distribution. Number of simulated caches was the same as in the observed data. (C) Principal component analysis of $p^{b i a s}$ values. Top: coefficients of the first two principal components for all birds. Red circles and numbers indicate birds shown in (A) and (B). Bottom: the first principal component. Birds cluster into center-preferring and edge-preferring groups. (D) Pearson correlation of $p^{b i a s}$ between subsets of all sessions paired within bird, between different birds, and between birds selected from the same PC1 cluster shown in (C). (E) Pearson correlation of $p^{b i a s}$ between pairs of trials at different trial lags. Values are medians across birds. Error bars: sem. Dashed line: linear regression. Gray shade: 95% confidence interval of linear regression fit to bootstrapped birds.

There are several possible explanations for the observed biases. One possibility is that biases are broadly shared across individuals of the species – akin to the preference for locations near walls in rodents (Lamprea et al., 2008). Another possibility is that biases are shared across some subgroups of individuals, but not all members of the species. Finally, biases could be unique to individual birds. To distinguish between these possibilities, we first asked whether there were common types of cache distributions across birds. We calculated principal components of all $p^{b i a s}$ values (Figure 2C). Remarkably, birds formed two clusters that were well-separated along the first principal component (Figure 2—figure supplement 1A). The first principal component resembled a bump in the center of the arena (Figure 2C), and accounted for 59% of the variance across individuals (Figure 2—figure supplement 1B). Thus, there were two groups of birds: those that preferred to cache in the center of the arena, and those that preferred the edges.

We next asked whether membership in one of these two groups was sufficient to explain the observed spatial biases. For each bird, we computed cache distributions separately for two randomly chosen, equal-sized subsets containing the same number of behavioral sessions. We then measured the Pearson correlation between the two distributions from the same bird (‘within-bird correlation’) and between distributions from different, randomly paired birds (‘across-bird correlation’). Within-bird correlation values were significantly higher than across-bird correlation values (p < 0.001, Figure 2D), even when birds were paired exclusively with birds from the same group (center-preferring or edge-preferring). Therefore, birds exhibited individual-specific biases that could not be entirely explained by common biases across a group of birds. An example of this is evident in Figure 2A: although birds 3 and 4 were both center-preferring birds, bird 3 had some additional bias for centrally-positioned rows and columns of the arena.

Finally, we asked whether spatial biases in individual birds were stable or drifted over time. We measured the Pearson correlation between cache distributions computed on pairs of sessions separated by different lags (Figure 2E). The slope of the relationship between lag and correlation was not significantly different from zero (p = 0.49, Bootstrap test), indicating that there was no detectable drift in bias.

We repeated all of the above analyses for the distribution of site checks in the Retrieval task. In this case, $p^{b i a s}$ was computed by dividing the number of times a particular site was checked by the total number of all site checks. All the results were similar to the ones obtained for caches in the Caching task (Figure 2—figure supplement 2). Collectively, our results suggest that chickadees exhibit idiosyncratic but stable biases, both in the sites that they check and those that they choose for caching. In both cases, the $p^{b i a s}$ distribution therefore serves as a valid baseline for subsequent models, described below.

Probabilistic models of chickadee behaviors

How do chickadees choose where to cache? One model consistent with our analysis so far is that birds choose each cache location by drawing it randomly from the distribution $p^{b i a s}$ . However, this is not the only possibility because $p^{b i a s}$ is computed by pooling across all time points. On a moment-by-moment basis, the likelihood of caching into a particular site may differ from the value given by $p^{b i a s}$ . For example, the choice of cache site could depend on the proximity of a site to the bird: chickadees may tend to cache close to the location they last interacted with or, conversely, they may avoid the vicinity of that location. The choice of a cache site may also depend on which sites are already occupied by seeds, since birds may tend to either cluster or to spread out their caches. The bird’s knowledge of which sites are occupied might be updated either by the act of caching, or by the act of checking the site to view its contents.

To model these possibilities, we considered three types of ‘special’ places in the arena at each moment in the Caching task: (1) the last site or feeder that the bird interacted with (‘previous site’), which includes checking a site or retrieving a seed, (2) sites that are currently occupied by cached seeds (‘occupied sites’), and (3) sites that have been previously checked by the bird and are currently empty (‘checked-empty sites’) (Figure 3A). We then constructed three scaling factors whose values changed with proximity to each of these special places (Figure 3B, defined in Materials and methods). The likelihood of caching into a particular site was computed by multiplying all three scaling factors by the corresponding value of $p^{b i a s}$ and normalizing the product to a sum of 1 across the arena. In other words, our model took each bird’s individual biases into account and then asked whether caching probabilities were additionally influenced, on a moment-by-moment basis, by special places in the arena.

Figure 3 with 3 supplements see all

Download asset Open asset

Schematic of the probabilistic behavioral models.

Models here compute the probability of caching into a particular site in the Caching task. Equivalent models are used for the probability of checking a particular site in the Retrieval task. (A) Schematic of the arena at an example time point in the Caching task depicting ‘special locations’ in the arena: previous site interaction, occupied sites, and checked-empty sites. (B) Examples of scaling factors that change with proximity to each of the special locations shown in (A). Probability distribution of caching into different sites is computed by multiplying all scaling factors by $p^{b i a s}$ . (C) Example models. For each model, some parameters are fixed at 0, while others are free and fit using maximum-likelihood estimation. For free parameters, examples shown here use the same values as in (B). All four models are plotted on the same color scale.

Each scaling factor was a Gaussian function across space, defined by two parameters. The first parameter was the amplitude of the Gaussian ( $γ_{p r v}$ , $γ_{o c c}$ , or $γ_{e m p}$ for the previous, occupied, and checked-empty sites, respectively), and indicated the strength and direction of the corresponding effect on probability, measured on a logarithmic base-10 scale. For example, $γ_{o c c} = - 1$ meant that the act of caching into a site decreased the likelihood of caching there again by a factor of 10. The second parameter was the width of the Gaussian ( $σ_{p r v}$ , $σ_{o c c}$ , or $σ_{e m p}$ , respectively) and indicated the spatial extent of each effect. In other words, a large value of $σ_{o c c}$ meant that the act of caching into a site affected the likelihood of subsequently caching into neighboring sites as well. This model was first applied to the data obtained in the Caching task. For each bird, we optimized the values of the model parameters in order to maximize the likelihood of the experimentally observed caches (Brunton et al., 2013; Raposo et al., 2012; Scott et al., 2017).

We also asked how chickadees chose which sites to check in the Retrieval task. Much like the choice of cache location, the choice of check location may be affected by proximity to the previous site, to occupied sites, and to checked-empty sites. We therefore used the same model as above, but fit to locations of checks instead of caches. At first, we fit the model only to the checks that the bird made up until and including finding a cache in the Retrieval task. We did this because after finding the first cache, birds often recached food, which confounded the analysis by mixing caching and retrieval behaviors in the same time period. We will analyze recaching separately later.

Our approach was to start with a baseline model (Model #0), in which the likelihood of choosing a particular site for caching or checking was given by $p^{b i a s}$ (Figure 3C). In this model, there were no free parameters, and the values of $γ_{p r v}$ , $γ_{o c c}$ , $γ_{e m p}$ , $σ_{p r v}$ , $σ_{o c c}$ , and $σ_{e m p}$ were all set to 0. We then gradually introduced free parameters to the model, one or two at a time. With each parameter introduction, we tested whether model performance was improved by measuring the Akaike Information Criterion (AIC), which quantifies model likelihood accounting for the number of free parameters (Dziak et al., 2020; Lubke et al., 2017).

Our model makes several assumptions. First, it uses a Gaussian function for each of the scaling factors. Second, it assumes a uniform spatial extent of this Gaussian across the arena. Finally, by multiplying scaling factors together it assumes that they have independent effects on probability. In separate models, we tested these assumptions and showed that chickadee behavior can be explained even better by including nonlinear interactions between scaling factors (Figure 3—figure supplement 1) by using functions that have more complex shapes than a Gaussian (Figure 3—figure supplement 2), and by allowing different spatial extents in different parts of the arena (Figure 3—figure supplement 3). However, these improvements were relatively modest and did not affect the main results described below.

Chickadee behaviors exhibit a strong proximity effect

We first asked how the location of the previous site affected the chickadee’s behavior. For this analysis we constructed Model #1, in which $γ_{p r v}$ and $σ_{p r v}$ were introduced as free parameters (Figure 4). We found that this model performed significantly better than Model #0 (see Table 1 for statistics of all model comparisons). The best-fit value of $γ_{p r v}$ was positive in all birds (0.94 ± 0.21), and the best-fit value of $σ_{p r v}$ was 15.5 ± 0.9 cm (see Table 2 for statistics of all model parameters). Thus, there was roughly a 10-fold (i.e., 10^0.94) increase in probability of caching next to the previous site, and the spatial extent of this effect was about 1/4 of the arena. To verify this result independently of the model, we compared the observed and expected distances of caches to previous sites and indeed found that birds cached closer to previous sites than expected by chance given each bird’s spatial bias (Figure 4—figure supplement 1).

Table 1

Summary of model performance.

Statistical analysis for all 12 models (0 through 11) presented in the Results section. Each model is compared to one of the previous models using bootstrap analysis of AIC values (see Materials and methods). p-values < 0.05 are emphasized in bold. Model #0 does not have corresponding p-values because it was the baseline model that was not compared to anything. Some models were only fit to data from the Caching or Retrieval task; in these cases, a hypen indicates that the corresponding fit was not performed.

Model #	Free parameters	Compared toModel #	Model improvement?
Model #	Free parameters	Compared toModel #	Caching task p-value	Retrieval task p-value
0	none	-	-	-
1	$γ_{p r v}; σ_{p r v}$	0	< 0.001	< 0.001
2	$γ_{p r v}; σ_{p r v}; γ_{o c c}$	1	< 0.001	< 0.001
3	$γ_{p r v}; σ_{p r v}; γ_{o c c}; γ_{e m p}$	2	< 0.001	0.015
4	$γ_{p r v}; σ_{p r v}; γ_{o c c}; γ_{e m p}; σ_{o c c}$	3	0.48	0.35
5	$γ_{p r v}; σ_{p r v}; γ_{o c c}; γ_{e m p}; σ_{e m p}$	3	0.93	0.42
6	$γ_{p r v}; σ_{p r v}; γ_{o c c}^{c}$	1	-	< 0.001
7	$γ_{p r v}; σ_{p r v}; γ_{o c c}^{c}; γ_{o c c}^{r}$	6	-	< 0.001
8	$γ_{p r v}; σ_{p r v}; γ_{o c c}; γ_{e m p}; τ_{o c c}$	3	0.48	-
9	$γ_{p r v}; σ_{p r v}; γ_{o c c}; γ_{e m p}; τ_{e m p}$	3	0.24	-
10	$γ_{p r v}; σ_{p r v}; γ_{o c c}; γ_{e m p}; ν_{o c c}$	3	0.97	-
11	$γ_{p r v}; σ_{p r v}; γ_{o c c}; γ_{e m p}; ν_{e m p}$	3	0.29	-

Figure 4 with 1 supplement see all

Download asset Open asset

Effect of proximity to the previous site in Caching and Retrieval tasks.

(A) Comparison of Model #1 to Model #0 applied to the Caching task. Here and in subsequent figures, ΔAIC indicates the difference in AIC between the model being considered (in this case Model #1, labeled on the x-axis) and the baseline model (in this case Model #0, labeled in the white box). Compared to Model #0, Model #1 uses two additional free parameters ( $γ_{p r v}$ and $σ_{p r v}$ ), specified in the colored box. Horizontal black line: ΔAIC value for caches pooled from all birds. Shaded area: distribution of 1000 ΔAIC values on data bootstrapped across birds. Values less than 0 indicate model improvement. Asterisk indicates statistically significant improvement. (B) Best-fit values of model parameters prv and σprv for Model #1 applied to the Caching task. Symbols indicate values for individual birds. Black line: median values across birds. (**C, D**) Same as (**A, B**), but for models applied to the Retrieval task.

Table 2

Contributions of different factors to behavior.

Statistical analysis of all $γ$ parameters mentioned in the Results section, which indicate the contribution of different factors to behavior. Best fit parameter values are indicated as median ± standard error of the median across all birds (N = 10 in the Caching task and N = 7 in the Retrieval task). Significance across birds is calculated using a one-sided Wilcoxon rank sum test to determine whether values across the population are significantly different from 0. The fraction of birds that are individually significant is calculated by bootstrapping the behavioral sessions of each bird. A p-value of 0.05 is used to assign significance in this case. See Materials and methods for details.

Parameter	Best-fit parameter value		Significance across birds		Fraction of individually significant birds
Parameter	Caching task	Retrieval task	Caching task p-value	Retrieval task p-value	Caching task	Retrieval task
$γ_{p r v}$	0.94 ± 0.21	1.83 ± 0.42	< 0.001	< 0.001	8/10	7/7
$γ_{o c c}$	–0.32 ± 0.17	0.26 ± 0.09	< 0.001	0.008	8/10	7/7
$γ_{e m p}$	0.13 ± 0.05	–0.14 ± 0.07	0.005	0.008	6/10	5/7
$γ_{o c c}$ ,first check	–0.28 ± 0.17	0.18 ± 0.10	0.005	0.008	8/10	2/7
$γ_{e m p}$ ,first check	0.12 ± 0.05	-	0.042	-	6/10	-
$γ_{o c c}^{c}$	-	0.42 ± 0.05	-	0.008	-	7/7
$γ_{o c c}^{r}$	-	0.26 ± 0.07	-	0.008	-	6/7

We applied the same model to site checks in the Retrieval task. Again, Model #1 performed better than Model #0. The value of $γ_{p r v}$ was positive in all birds (1.83 ± 0.42), indicating a tendency to check sites close to the previous sites, and $σ_{p r v}$ was 12.9 ± 1.7 cm. This result is consistent with our qualitative observations of chickadees in the Retrieval task: as birds moved through the arena – sometimes in the direction of a hidden cache – they often checked sites on their path. In fact, the value of $γ_{p r v}$ was significantly higher in the Retrieval task (p < 0.05, unpaired t-test), but the best-fit $σ_{p r v}$ was not significantly different between the two tasks (p = 0.54, unpaired t-test). These results show that birds are influenced by proximity during retrieval more than during caching, even though the spatial extent of this effect is similar during the two behaviors.

Chickadee behaviors are affected by site content

We next asked whether chickadee behaviors were additionally influenced by the contents of individual sites (Figure 5A–B). In the Caching task, we fit Model #2, which included $γ_{o c c}$ as a third free parameter to model the effect of occupied sites. Model #2 performed even better than Model #1, indicating that occupied sites indeed affected the behavior. We then fit Model #3, in which $γ_{e m p}$ was additionally introduced as a fourth free parameter to model the effect of checked-empty sites. Model #3 performed better yet, indicating that checked-empty sites also affected the behavior. Across birds, the value of $γ_{o c c}$ was negative (–0.32 ± 0.17), whereas the value of $γ_{e m p}$ was positive (0.13 ± 0.05). These values indicate roughly a twofold decrease in probability of caching into occupied sites and a 1.3-fold increase in probability of caching into checked-empty sites. Both results are consistent with chickadees spreading their caches throughout the arena, rather than pooling them into the same sites. Indeed, chickadees tended to cache into occupied sites less often than expected by chance (Figure 5—figure supplement 1A).

Figure 5 with 1 supplement see all

Download asset Open asset

Site-specific, opposing effects of site content on behavior in Caching and Retrieving tasks.

(A) Performance of Models #2 and #3 applied to the Caching task, plotted as in Figure 4A. These models introduce free parameters to quantify the effect of occupied and checked-empty sites on behavior. (B) Best-fit values of $γ_{o c c}$ and $γ_{e m p}$ applied to the Caching task, plotted as in Figure 4B. (**C, D**) Same as (**A, B**) but for models applied to the Retrieval task. (E) Performance of Models #4 and #5 applied to the Caching task, plotted as in Figure 4A. These models introduce free parameters to quantify the spatial extent of the effect of occupied and checked-empty sites on behavior. (F) Best-fit values of $σ_{o c c}$ and $σ_{e m p}$ applied to individual birds in the Caching task, plotted as in Figure 4B. (**G, H**) Same as (**E, F**), but for models applied to the Retrieval task.

We again applied the same models to site checks in the Retrieval task (Figure 5C–D). As in the Caching task, introducing $γ_{o c c}$ and $γ_{e m p}$ significantly improved the fit of the model. However, the signs of the two parameters were reversed compared to the Caching task: $γ_{o c c}$ was positive in all birds (0.26 ± 0.09), whereas $γ_{e m p}$ was negative in all birds (–0.14 ± 0.07). These results show that birds were attracted to occupied sites during the feeder-closed phase of the task. This is consistent with chickadees searching for their caches. Indeed, chickadees found their caches after fewer checks than expected by chance: 8.5 ± 5.6 checks in the observed data, compared to 20.8 ± 4.5 checks in data where trajectories were shuffled between trials (Figure 5—figure supplement 1B).

Our analyses so far show that chickadees can be attracted to or avoid specific sites depending on the content of these sites. We asked whether these behavioral effects were site-specific, or whether neighboring locations were affected as well. We introduced either $σ_{o c c}$ or $σ_{e m p}$ as an additional free parameter (Model #4 and Model #5 respectively, Figure 5E–H). Neither model was an improvement over Model #3. In fact, the best-fit values of $σ_{o c c}$ and $σ_{e m p}$ in both models were between 0.01 and 0.02 cm, much smaller than the distance between neighboring sites (7.1 cm). Thus, the effects of site contents on chickadee behavior were site-specific, and were small even at neighboring sites of the arena. We verified this result independently of the model by measuring the fraction of caches and checks at various distances away from occupied sites (Figure 5—figure supplement 1C–D). This site specificity is consistent with that demonstrated under different conditions by previous work (Cowie et al., 1981).

Chickadees have memories of site contents

We have shown that chickadees avoid caching into sites that are occupied, but increase their probability of caching into checked-empty sites. An intriguing explanation of these results is that chickadees use memory of site contents to influence their caching behavior. However, after obtaining a seed, a chickadee may simply check multiple sites in the arena and preferentially cache into an empty one that it encounters. This would be a visually-guided strategy that does not require memory. We considered this possibility unlikely because chickadees usually cached into the first site that they opened after visiting the food source (79% of caches). However, we wanted to ensure that the remaining 21% of caches were not responsible for the observed effects. We implemented the same models as before (Models #2 and #3), but instead of fitting them to the observed caches, we fit them to the locations of the first sites opened by the bird after obtaining a seed (‘first check’). We found that $γ_{o c c}$ in Model #2 was still significantly negative, and $γ_{e m p}$ in Model #3 was still significantly positive (Figure 6A–B). Therefore, chickadees did not merely check sites until finding an empty one. Rather, they actually used memories of which sites were occupied and which were empty.

Figure 6 with 2 supplements see all

Download asset Open asset

Results are best explained by long-lasting memories.

(**A, B**) Values of parameters across birds, compared between models applied to caches or first checks in the Caching task. Values of $γ_{o c c}$ are from Model #2, and values of $γ_{e m p}$ are from Model #3. (C) Values of $γ_{o c c}$ across birds, compared between Model #2 applied to checks or the same model applied to first checks in the Retrieval task. (D) Performance of Models #6 and #7 applied to the Retrieval task, plotted as in Figure 4A. These models introduce free parameters to separately quantify the effects of caches and recaches on the site checking behavior. (E) Best-fit values of $γ_{o c c}^{c}$ and $γ_{o c c}^{r}$ applied to individual birds in the Retrieval task, plotted as in Figure 4B. (F) Performance of Models #8 and #9 applied to the Caching task, plotted as in Figure 4A. These models introduce free parameters $τ_{o c c}$ and $τ_{e m p}$ to quantify the decay of $γ_{o c c}$ and $γ_{e m p}$ over time since the site was last checked by the bird. (G) Best-fit values of $τ_{o c c}$ and $τ_{e m p}$ applied to individual birds in the Caching task, plotted as in Figure 4B.

We also fit Model #2 to first checks in the Retrieval task. In this case, $γ_{o c c}$ was still positive. In other words, chickadees had an increased probability of checking an occupied site even on the first attempt. This implies that chickadees did not search the arena until finding a cache, but used memories of occupied sites. Note that $γ_{e m p}$ was not computed for this analysis because, by definition, there could not be any checked-empty sites before the first check.

Finally, we wanted to make sure that chickadees did not use olfactory cues to detect seeds in cache sites. Chickadees are not known to use olfaction for finding food (Cowie et al., 1981; Herz et al., 1994; Sherry, 1984a), and previous studies have excluded this possibility by removing caches during the delay period of the task. We repeated this experiment in our arena. Consistent with published results, even when caches were removed from the arena, chickadees tended to spend more time at the locations of the missing caches after a delay period of up to 1 hr (Figure 6—figure supplement 1). Collectively, our results show that chickadees have memories of site contents, as in previous work (Clayton and Dickinson, 1999; Sherry, 1984a). They use these memory of site contents in flexible ways, both during caching and during retrieval.

Memories of site contents last the duration of the session

We next asked for how long memories lasted in our arena. In the feeder-closed phase of the Retrieval task, chickadees often ‘recached’ seeds after retrieving them (44% of all retrievals). Therefore, there were two types of caches in this task: recent recaches and older caches made prior to the delay period. We first analyzed how these different types of caches affected behavior. For this analysis, we introduced models with separate parameters $γ_{o c c}^{c}$ and $γ_{o c c}^{r}$ instead of $γ_{o c c}$ to fit the contribution of older caches and recaches, respectively. These models were fit to the entire feeder-closed phase of the Retrieval task, which contained both the retrieval and the recaching behaviors. Model #6 introduced $γ_{o c c}^{c}$ as a free parameter, while Model #7 introduced both $γ_{o c c}^{c}$ and $γ_{o c c}^{r}$ . We found that Model #6 fit the data significantly better than Model #1, and Model #7 fit the data even better than Model #6 (Figure 6D). The best-fit values were significantly positive for both parameters, but higher for $γ_{o c c}^{r}$ than for $γ_{o c c}^{c}$ (0.42 ± 0.05 vs 0.26 ± 0.07, respectively, p < 0.01 for the comparison). These results indicate that birds were attracted to all of their caches during retrieval, but had a stronger preference for the recent recaches.

One possible explanation of this result is that birds had a memory decay that reduced the effect of older caches on behavior. However, birds often recache a single seed multiple times, and may simply have a preference for revisiting recache locations. We therefore performed a different analysis that modeled a continuous memory decay. The Caching task was ideally suited for this analysis because it did not have a trial structure; caches could be retrieved at any later time in the session. We defined a ‘memory decay factor’ for each site, which was reset to one whenever that site was checked and decayed exponentially to 0 afterwards. In Model #8, this factor was multiplied by $γ_{o c c}$ and decayed with a timescale $τ_{o c c}$ . In Model #9, this factor was multiplied by $γ_{e m p}$ and decayed with a timescale $τ_{e m p}$ . Both models were compared to Model #3, in which neither $γ_{o c c}$ nor $γ_{e m p}$ decayed. We found that neither model improved the fit to the data (Figure 6F). The best-fit values of $τ_{o c c}$ and $τ_{e m p}$ were very large (both >44 min, Figure 6G). Therefore, there did not appear to be a significant memory decay at the timescale of our experiments. This additionally implies that the preference for retrieving recaches may be a behavioral preference unrelated to a memory decay.

Chickadees often recheck sites, with a median of 2.4 min between consecutive checks of the same site. Our results suggest that these rechecks are not necessary for maintaining memory, since the memory decay in the above analysis is reset at every site check. To verify this independently of the model, we measured the fraction of caches that were made into empty sites. In this analysis, we only included sites that had not been recently checked by the bird. We found that chickadees prefer to cache into empty sites even after not checking sites for a long period of time (p < 0.001 after 30 min; Figure 6—figure supplement 2). Our results do not indicate whether memory remains stable over periods much longer than the behavioral session (1 hr), and whether rechecks become necessary at those timescales. However, prior work has showed some memories lasting up to a month, even when birds are not allowed to recheck their sites (Roth et al., 2012). Additionally even 1 hr delays are ethologically important for chickadees, since they typically retrieve their caches over the course of a single day in the wild (Cowie et al., 1981).

Memories of site contents are of high-capacity and accessed in an arbitrary order

Food-caching birds are famous for storing and likely remembering large quantities of food items in the wild (Pravosudov, 1985). So far, it is unclear if chickadees exhibit a high memory capacity in our arena. For example, the statistical effects we observed can be explained by chickadees remembering a small number of caches and having no memory of other caches. To test this possibility, we allowed $γ_{o c c}$ and $γ_{e m p}$ to decay exponentially with the total number of occupied sites and checked-empty sites in the arena. The decay constants of these exponentials were two additional parameters, $ν_{o c c}$ and $ν_{e m p}$ , respectively. As before, we fit two models (Model #10 and Model #11), each introducing one of these two parameters. Neither model fit the data better than Model #3, in which $γ_{o c c}$ and $γ_{e m p}$ did not decay (Figure 7A). The best-fit values of $ν_{o c c}$ and $ν_{e m p}$ were >15 sites in both cases (Figure 7B) – comparable to the maximum number caches present in the arena at any single moment in the session (15.0 ± 1.0, N = 101 sessions). Therefore, memory did not appear to decay at capacities tested by our experiments.

Figure 7

Download asset Open asset

Memories are high capacity, and accessed in arbitrary order.

(A) Performance of Models #10 and #11 applied to the Caching task, plotted as in Figure 4A. These models introduce free parameters $v_{o c c}$ and $v_{e m p}$ to quantify the decay of $γ_{o c c}$ and $γ_{e m p}$ as either the number of occupied sites or the number of checked-empty sites in the arena increases, respectively. (B) Best-fit values of $v_{o c c}$ and $v_{e m p}$ applied to individual birds in the Caching task, plotted as in Figure 4B. (C) Fraction of retrievals as a function of order in which seeds were caches. Order is aligned to the first cache. Red marker: mean observed fraction across birds. Black line: average value expected from retrieving seeds in a random order. Grey: 95% confidence interval of retrievals made in a random order. (D) Same as (C), but aligned to the most recent cache. Arrow indicates the only point that was significantly outside the 95% confidence interval of shuffled data. (E) Same as (D), but with transient caches eliminated. Transient caches are those that are retrieved without leaving the perch.

All of our models so far have used spatial location to explain the choice of sites for caching and checking. We also asked whether the order in which seeds were cached was predictive of the order in which birds chose to retrieve them. For this analysis, we used the Caching task, in which birds often cached and retrieved large numbers of seeds. Across all retrievals, we asked how often the nth cached seed was chosen by the bird (Figure 7C). There was no difference between observed data and shuffled data, in which seeds were chosen randomly from the available caches. Thus, there was no detectable ‘primacy’ effect in the data (p = 0.96). We then asked how often the nth most recent seed was chosen. In this case, the most recently cached seed was retrieved more often than expected by chance (p < 0.001, Figure 7D). However, this ‘recency’ effect was largely caused by caches that the bird retrieved without leaving the site of the cache, which accounted for 14% of all retrievals (Figure 7E). We often observed chickadees making several of these ‘transient caches’ and retrievals with a single seed, before finally settling on a cache location. Once transient caches were excluded from the data, there was no detectable primacy or recency effect in the order of retrieval (p = 0.44 and p = 0.24, respectively).

Discussion

We designed an apparatus to explore how chickadees make navigational decisions during food caching and retrieval. We built quantitative models to understand the relative contributions of mnemonic and non-mnemonic strategies to these behaviors. Our setup allowed birds to move within an arena similar to those used for rodent experiments (Lever et al., 2002; Muller et al., 1987; Wilson and McNaughton, 1993)^. We found that several natural features of chickadee spatial navigation were preserved in this environment. Our apparatus also allowed birds to cache and retrieve food from a relatively large number of concealed sites within the arena. Tracking behavior at these sites revealed that chickadees rely on multiple strategies for caching and retrieval and use memory in flexible ways during these behaviors.

We found two spatial strategies that did not require birds to remember the contents of individual sites: the use of spatial biases and a proximity effect. First, chickadees had spatial biases that were stable over time and unique to individual birds. Previous work has shown spatial biases in the wild. For example, individual birds can cache in different zones within a shared territory or on different parts of a tree (Cowie et al., 1981; Dally et al., 2006; Lahti et al., 1998). This strategy has a clear ethological advantage: biases reduce the memory load of cache locations, whereas choosing biases idiosyncratically minimizes unwanted overlap between individuals. The second contribution to behavior was the effect of proximity: chickadees tended to cache and check sites close to their previous site interaction. These effects have also been observed in the wild and have ethological advantages. Caching close to the food source minimizes travel distance and exposure to dangers (Clarkson et al., 1986; Scott et al., 2017; Sherry et al., 2010; Stapanian and Smith, 1978)^. Checking locations in close proximity is an efficient opportunistic method of exploring the environment (Krebs et al., 1974; van der Vaart et al., 2011). Although the patterns we find are on a dramatically smaller scale, it is intriguing to speculate that biases and proximity effects are driven by shared mechanisms across conditions.

We do not know whether these effects are truly non-mnemonic on timescales longer than our behavioral sessions. For example, a bird’s experience over long periods of time may influence its biases or its typical trajectories within an environment (Hampton and Sherry, 1994; Raby et al., 2007). Regardless, explaining away biases and proximity effects was critical for isolating the remaining memory-guided components in our experiments.

Our analysis has revealed several features of chickadee memory. Memories were spatially specific, lasted the duration of the hour-long session, had high capacity, and could be retrieved in an order different from their storage order. These features have been studied before – usually in experiments specifically designed to look at each feature of memory in isolation. For example, radioactive tagging of food in the wild has shown that birds search for caches with centimeter precision (Cowie et al., 1981)^. In both field and laboratory experiments, birds continue to remember locations on both hour-long timescales (Clayton and Krebs, 1994b; Cowie et al., 1981; Sherry and Vaccarino, 1989) and at least a month later (Ekman et al., 1996; Roth et al., 2012). Chickadees can cache thousands of seeds per day and are suspected to remember a large fraction of them (Brodin, 2010; Pravosudov, 1985)^. Finally, previous studies have also not detected a relationship between the order of caching and the order of retrieval, suggesting that chickadees have independent memories of caches rather than a single memory of the action sequence (Balda and Kamil, 1989; Sherry, 1984a; Sherry, 1984b). Our models capture these features on smaller spatial and temporal scales, and notably combine them within a single laboratory paradigm.

Classic work has shown that food-caching birds use memory to retrieve food (Krushinskaya, 1966; Sherry, 1984a; Clayton and Dickinson, 1998). They can even remember the contents of individual sites – for example, to choose caches that contain a certain type of food (Clayton and Dickinson, 1998; Sherry, 1984a). We also observed memories of site contents in our arena. Chickadees were attracted to occupied sites during retrieval, but avoided empty sites that they had previously checked. These results show that memories of site contents in chickadees not only affect a single behavior, but even influence the choice between different behaviors – in our case, site attraction or avoidance.

Site content influenced not only retrieval, but also caching itself. When caching, birds avoided occupied sites, but preferred sites that they had checked and found to be empty. These behaviors resulted in a ‘spreading’ of caches throughout an environment. Cache spreading is observed on a larger scale in the wild and is thought to be a defensive strategy against pilferers (Alpern et al., 2012; Cowie et al., 1981). However, it was not previously known that memory plays a role in this behavior. The implication of our results is that a single memory of a cached seed can be used by the bird for entirely different behaviors depending on context – in our case, caching or retrieval.

It has been shown that food-caching birds use memory in flexible ways. For example, they choose caches with certain types of food instead of others depending on selective satiety (Clayton and Dickinson, 1999) or knowledge about food perishability (Clayton and Dickinson, 1998; Feeney et al., 2011). Our results further show that memories might even be used for actions that are unrelated to immediate feeding needs – in our case to spread caches throughout an environment. The ability of the same memory to flexibly drive different behaviors that have different goals is a hallmark of episodic memories in humans (Allen and Fortin, 2013; Clayton et al., 2003; Tolman, 1948; Tulving, 1972)^. Our findings lends support to the idea that memories of food-caching birds are similarly general-purpose in nature.

Model index $(M)$	Parameter												# free param $(f_{M})$
	$γ_{p r v}$	$σ_{p r v}$	$γ_{o c c}$	$σ_{o c c}$	$γ_{e m p}$	$σ_{e m p}$	$γ_{o c c}^{c}$	$γ_{o c c}^{r}$	$1 / τ_{o c c}$	$1 / τ_{e m p}$	$1 / ν_{o c c}$	$1 / ν_{e m p}$
	Parameter index ( $z$ )
	1	2	3	4	5	6	7	8	9	10	11	12
0	{0}	{0}	{0}	{0}	{0}	{0}	{0}	{0}	{0}	{0}	{0}	{0}	0
1	(-∞,∞)	[0,∞)	{0}	{0}	{0}	{0}	{0}	{0}	{0}	{0}	{0}	{0}	2
2	(-∞,∞)	[0,∞)	(-∞,∞)	{0}	{0}	{0}	{0}	{0}	{0}	{0}	{0}	{0}	3
3	(-∞,∞)	[0,∞)	(-∞,∞)	{0}	(-∞,∞)	{0}	{0}	{0}	{0}	{0}	{0}	{0}	4
4	(-∞,∞)	[0,∞)	(-∞,∞)	[0,∞)	(-∞,∞)	{0}	{0}	{0}	{0}	{0}	{0}	{0}	5
5	(-∞,∞)	[0,∞)	(-∞,∞)	{0}	(-∞,∞)	[0,∞)	{0}	{0}	{0}	{0}	{0}	{0}	5
6	(-∞,∞)	[0,∞)	{0}	{0}	{0}	{0}	(-∞,∞)	{0}	{0}	{0}	{0}	{0}	3
7	(-∞,∞)	[0,∞)	{0}	{0}	{0}	{0}	(-∞,∞)	(-∞,∞)	{0}	{0}	{0}	{0}	4
8	(-∞,∞)	[0,∞)	(-∞,∞)	{0}	(-∞,∞)	{0}	{0}	{0}	[0,∞)	{0}	{0}	{0}	5
9	(-∞,∞)	[0,∞)	(-∞,∞)	{0}	(-∞,∞)	{0}	{0}	{0}	{0}	[0,∞)	{0}	{0}	5
10	(-∞,∞)	[0,∞)	(-∞,∞)	{0}	(-∞,∞)	{0}	{0}	{0}	{0}	{0}	[0,∞)	{0}	5
11	(-∞,∞)	[0,∞)	(-∞,∞)	{0}	(-∞,∞)	{0}	{0}	{0}	{0}	{0}	{0}	[0,∞)	5

Share this article

Cite this article

Behavioral paradigm for food caching and retrieval in chickadees.

Chickadees exhibit idiosyncratic and stable biases in the locations of caches.

Schematic of the probabilistic behavioral models.

Summary of model performance.

Effect of proximity to the previous site in Caching and Retrieval tasks.

Contributions of different factors to behavior.

Site-specific, opposing effects of site content on behavior in Caching and Retrieving tasks.

Results are best explained by long-lasting memories.

Memories are high capacity, and accessed in arbitrary order.

Masks for different subsets of interactions.

Permitted ranges of parameters, P(M,z), for main models.

Author details

Marissa C Applegate

Contribution

Competing interests

Dmitriy Aronov

Contribution

For correspondence

Competing interests

Downloads (link to download the article as PDF)

Open citations (links to open the citations from this article in various online reference manager services)

Cite this article (links to download the citations from this article in formats compatible with various reference manager tools)

Categories and tags

Further reading

Permitted ranges of parameters, $P (M, z)$ , for main models.