Get started
Work

Author response: Flexible use of memory by food-caching birds

Year: 2021
Type: peer-review
Abstract: Article Figures and data Abstract Editor's evaluation eLife digest Introduction Results Discussion Materials and methods Data availability References Decision letter Author response 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 Decision letter eLife's review process 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 (pbias) 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 pbias 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 pbias 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 pbias 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 pbias 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 pbias 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 pbias 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, pbias 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 pbias 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 pbias. However, this is not the only possibility because pbias 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 pbias. 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 values with proximity to each of these special places (Figure in Materials and methods). The likelihood of caching into a particular site was computed by all three scaling factors by the corresponding value of pbias and the to a of 1 across the arena. In other our model each bird’s individual biases into and then asked whether caching were on a moment-by-moment basis, by special places in the arena. Figure 3 with 3 supplements see all Download asset Open asset of the probabilistic behavioral the of caching into a particular site in the Caching task. models are used for the of checking a particular site in the Retrieval task. (A) of the arena at an example time in the Caching task in the site occupied sites, and sites. (B) of scaling factors that with proximity to each of the special locations shown in (A). Probability distribution of caching into different sites is computed by all scaling factors by pbias. (C) For each some are at while are free and fit using For free shown use the same values as in (B). All four models are on the same scaling was a across by two The first was the of the or for the and sites, and the and direction of the corresponding on measured on a For example, that the act of caching into a site the likelihood of caching there by a of The was the of the or and the spatial of each In other a large value of that the act of caching into a site the likelihood of caching into sites as This model was first applied to the data obtained in the Caching task. For each bird, we the values of the model in order to the likelihood of the observed caches (Brunton et al., 2013; Raposo et al., 2012; Scott et al., 2017). We also asked how chickadees which sites to check in the Retrieval task. like the choice of cache location, the choice of check location may be by proximity to the site, to occupied sites, and to sites. We therefore used the same model as but fit to locations of checks instead of caches. At we fit the model only to the checks that the bird and including a cache in the Retrieval task. We this because after the first birds food, which the analysis by caching and retrieval behaviors in the same time We separately Our approach was to with a baseline model in which the likelihood of a particular site for caching or checking was given by pbias (Figure In this there were no free and the values of and were all to We then free to the one or two at a time. each we whether model was by measuring the which model likelihood for the number of free et al., et al., 2017). Our model several it a for each of the scaling factors. it a uniform spatial of this across the arena. Finally, by scaling factors it that they have independent on In models, we these and that chickadee behavior can be explained even by including interactions between scaling factors (Figure supplement by using that have more than a (Figure supplement 2), and by allowing different spatial in different of the arena (Figure supplement However, these were and not the results described below. behaviors exhibit a proximity We first asked how the location of the site the chickadee’s behavior. For this analysis we constructed in which and were as free (Figure We found that this model performed significantly than 1 for of all model The value of was in all birds and the value of was cm 2 for of all model Thus, there was a in of caching next to the site, and the spatial of this was about of the arena. To this of the we compared the observed and expected of caches to sites and found that birds cached to sites than expected by chance given each bird’s spatial bias (Figure supplement 1). 1 of model analysis for all models through in the Results section. model is compared to one of the models using analysis of values Materials and methods). < are in does not have corresponding because it was the baseline model that was not compared to models were only fit to data from the Caching or Retrieval in these cases, a indicates that the corresponding fit was not task task Figure 4 with 1 supplement see all Download asset Open asset of proximity to the site in Caching and Retrieval (A) of to applied to the Caching task. and in subsequent indicates the in between the model considered this on the and the baseline model this in the to two additional free and in the line: value for caches from all birds. distribution of values on data bootstrapped across birds. Values than indicate model indicates (B) values of model and for applied to the Caching task. indicate values for individual birds. Black line: values across birds. Same as but for models applied to the Retrieval task. 2 of different factors to behavior. analysis of all in the Results which indicate the contribution of different factors to behavior. fit values are as of the across all birds = 10 in the Caching task and = in the Retrieval across birds is calculated using a to determine whether values across the are significantly different from The of birds that are individually is calculated by the behavioral sessions of each bird. A of is used to in this Materials and methods for across of individually task task We applied the same model to site checks in the Retrieval task. performed than The value of was in all birds indicating a to check sites close to the sites, and was This is consistent with our of chickadees in the Retrieval task: as birds through the arena – in the direction of a hidden cache – they checked sites on their In fact, the value of was significantly higher in the Retrieval task (p < but the was not significantly different between the two tasks (p = These results show that birds are by proximity during retrieval more than during caching, even though the spatial of this is similar during the two behaviors are by site We next asked whether chickadee behaviors were by the contents of individual sites (Figure In the Caching task, we fit which included as a free to model the of occupied sites. performed even than indicating that occupied sites the behavior. We then fit in which was as a free
Source: eLife
DOI: 10.7554/elife.70600.sa2
Language: English
Open Access status: gold
Estimated APC paid: $2,500
OpenAlex ID: W4226410801
Cites:
Cited by:
Related to: 10