Session A22: Animal Behavior

Uncovering the dynamical structure of behavioral repertoires

Animal behavior consists of an intricate hierarchy of dynamics, from brief muscle twitches to stereotyped behaviors to longer-lived states like hunger, aggression, and parenting. How does an animal bridge these timescales to create complex sequences of actions? The approach that most researchers take when studying sequences of behaviors tends to be probabilistic, observing how discrete states transition in a largely memoryless fashion. In this talk, we take a different approach, fitting dynamical models to long behavioral sequences from fruit flies and other species. We show that these models replicate many summary statistics of the underlying behavioral sequence data and that their fixed points have a geometry that mirrors the geometry of the animals’ behavioral repertoires. In addition, we show that the long timescales generated by this model are best explained by a hierarchy of interacting dynamical subsystems that is comparable to the hierarchical structure of behavior. These results suggest a new framework for uncovering the hidden states that modulate the behaviors that an animal performs and predicting how physiology may be linked to behavior.   

Collective bumblebee foraging in a controlled stochastic environment

We study bumblebees in a flight chamber with 10 artificial flowers to explore and accurately quantify foraging behavior. An RFID system is employed to identify the individual foragers and detect their presence in flowers as well as entry and exit of the hive. Flowers release rewards upon bee’s visit with a predefined probability and are divided into two groups of separate reward probabilities and color. Furthermore, we impose a reward refractory period, that is a time span that a bee needs to stay away from a recently triggered flower before it is able to trigger it again.
In this study, an entire colony of tagged bees is kept in the foraging chamber for 4-6 days. We study 7 colonies with 40-70 bees each. We analyze and quantify the operant learning of the relation between flower color and reward probability: bees are able to identify the more rewarding flower type, effectively matching the reward probability ratios. However, the number of bees who are able to do so strongly depends on the length of the imposed refractory period.
We conclude that a combined setting of refractory time and reward probability is too challenging for most bumblebees to learn in the given time span and thus leads to suboptimal foraging choices whereas they quickly learn to adapt to Bernoulli trials alone   

A Low-Cost Modular Camera System for 3D Pose Estimation in the Field

The individual behavior of animals determines the outcome of ecological interactions and drives group organization and community dynamics. Considerable attention has been given to understanding these types of interactions by way of automated tracking of individuals and bio-logging; however, these methods rarely investigate the behaviors performed by animals. Recent developments in deep learning for pose estimation provide a promising avenue for understanding individual behavior at resolutions previously not possible. However, systems for generating datasets appropriate for these methods in the field are lacking. To fill this gap, we developed a low-cost, modular camera platform to generate 3D imaging data compatible with contemporary pose estimation techniques. We use a connected network of solar-powered cameras that supports synchronous capture, triggering, and the integration of multisensory metadata. We initially generated datasets from three camera modules that allow for 3D pose estimation in a large outdoor area of over 600m² and demonstrate its use in an open field experiment at the Mpala Research Centre and Wildlife Foundation in Nanyuki, Kenya. Our system is designed to be modular and extensible, facilitating the use of many camera modules in very large open areas.   

Bridging time scales in C. elegans behavior through transfer operators

Animal behavior is modulated over multiple timescales: from fast control by neural activity to slower variation due to starvation or aging. Can we extract these hidden processes from the movement dynamics alone? We introduce a principled method to simultaneously reconstruct and partition the dynamical state-space, which we use to approximate the Perron-Frobenius operator. Our operator approximation is built to be maximally predictive and Markovian and its spectral decomposition provides a hierarchy of modes, which evolve over multiple time-scales. Applied to C. elegans locomotion, we find coherent structures which represent behavioral motifs, while the dynamics of the long-lived modes capture slower changes in behavioral “strategies”, e.g. the worm’s exploratory propensity. By subsuming the nonlinear dynamics into the process of partitioning and state space reconstruction, we obtain a model of the dynamics which is simple yet still able to faithfully reproduce the complexity of worm behavior from milliseconds to hours.   

Inferring behavioral homologies from dynamical models

Linking the evolution of animal behavior to the genes underlying it has proven challenging, largely due to our inability to find representations of behavior that allow for inter-species comparisons. Animals exhibit variability in many different traits across species, but certain traits are relatively conserved. These traits are known as homologies or homologous structures, and quantitatively identifying these homologies in behavior could provide a new approach for understanding the evolution of behavior. Here, we measure the behavioral repertoires of six fruit fly species, finding both the frequency of behavioral performance, as well as their temporal dynamics. By fitting dynamical models to these transitions, we can reproduce the summary statistics of our dataset, including long timescale dynamics and hierarchical structure. We show that features of these models can be used to define such homologies, providing future avenues for exploring the genetic basis of behavioral evolution.   

Long timescale dynamics in freely behaving rats

Natural behavior is composed of rich postural dynamics that contain stereotyped movements performed by the animal. These behaviors span multiple timescales and are performed in a structured manner during spontaneous behavior. Thus, a quantitative understanding of behavioral dynamics is crucial to help unravel the latent physiological states driving behavior. Here, we extract postural information from videos of freely moving rats in an arena using markerless tracking tools. Using these postural time series’ we create a low-dimensional behavioral state space using unsupervised methods that characterizes stereotypic behavioral bouts. We find long, non-Markovian timescales of predictability across novel and familiar trials of light and dark conditions in the arena. These behavioral sequences are found to be arranged in hierarchical clusters, similar to previous results in fruit flies. These results support hierarchical organization of behavior as a general principle across species for generating long timescale dynamics.   

Measuring and modeling the dynamics of the thermal memory of C. elegans

The roundworm C. elegans learns from its experiences. When placed on a thermal gradient, worms perform thermotaxis to or away from the conditioned temperature, depending on food abundance during the conditioning phase. To quantify this behavior, we developed a novel assay that tracks single worms—each experiencing a Spatio-temporal thermal gradient with thermal precision ±0.01C—in a small (2.8ul) droplets of buffer, arrayed on hydrophobic-printed microscope slides. CCD cameras monitor many worms simultaneously in many droplets, each droplet at 20C midpoint with a thermal gradient of 0.5C/cm. A worm’s thermal preference is summarized as a thermotaxis index, and the index dynamics are tracked at a high temporal resolution for many hours. Initially, worms reared at 15C and 25C exhibit cryophilic or thermophilic tendencies, respectively, and starvation during conditioning or in the droplet reverses these tendencies. This reversal is non-monotonic indicating multiple dynamic processes for learning that operate on different time scales. We build a predictive model with multiple time scales and utilize mutants to detangle the various learning processes. The model predicts the behavior under various conditions.   

Multi-animal pose tracking using deep neural networks

Dissecting behavior in freely moving animals at the fast timescales requires rich representations of their motor dynamics. Recently, we developed a method to automate the estimation of animal pose from videos using deep neural networks (Pereira et al., 2019). This method, termed LEAP, detects body part positions in single animal videos. Extending these techniques to a multi-animal context presents technical challenges, such as assigning body part positions to the correct animal. Here we present a new framework we term SLEAP (Social LEAP Estimates Animal Poses) that can explicitly model the relationship between body parts, enabling accurate multi-animal pose estimation. The framework implements multiple neural network meta-architectures which we empirically evaluate on tracking sub-tasks. We demonstrate the generalizability of this framework by applying this technique to videos of a variety of animals, including a high-resolution dataset of freely interacting fruit flies to construct a map of postural dynamics during courtship.   

Phenotype to Function: Predicting drug mode of action from behavioural fingerprints

Pesticides and anthelmintics (nematode-killing drugs) are discovered through phenotypic screens in target species. This means that their efficacy is often known early in the development pipeline, but their mode of action is not. Therefore, an important problem in developing new compounds to combat the rise of anthelmintic resistance is determining their mode of action. We record multiple worms from 96-well plates using a multi-camera imaging system and extract behavioural features from tracked animals to define a quantitative phenotype for each well in response to a library of ~80 drugs from 10 known mode of action classes. Because drug dose can have a strong effect on behavioural response, we record worms' response to a range of doses. We combine information across doses using multiple-instance learning to predict a compound's mode of action on unseen data.   

Probing the neural substrates of movement generation across the rodent behavioral repertoire

Mammals are generalists, capable of flexibly deploying movements across a range of behavioral contexts. However, existing paradigms for studying the genesis of movement in the brain often probe only a narrow range of highly trained behaviors, leaving the neural mechanisms that support this flexibility unclear. To extend the range of behaviors and contexts that can be studied, we developed a new behavioral monitoring system, CAPTURE, that combines motion capture and deep learning to track the 3D movements of twenty points on a freely behaving rat's trunk and appendages, continuously over week-long timescales. We combined CAPTURE with continuous neural recordings in the dorsolateral striatum, a brain region with a known, if debated, role in controlling diverse aspects of movement and behavior. Within individual behavioral states, striatal neurons displayed tuning to kinematic and behavioral variables. However across states, these tuning properties changed, in a manner that improved the within-state behavioral decodability. This suggests that context-specific coding may be a means to efficiently represent behavioral variables that are flexibly deployed across states, and that representational findings observed in tasks may fail to generalize.   

Speed selection and sampling strategies for terrestrial trail tracking

Terrestrial animals such as ants, mice and dogs use surface-borne odor trails to establish navigation routes or to find food and mates by following adsorbed chemical traces. Trail-tracking behavior is commonly observed, yet the strategies animals use to track trails are largely unknown. We examine generic features of trail-tracking by posing the problem as a search task of finding the trail after each loss of contact. We show that trail geometry imposes strong constraints on tracking speed; the maximal speed scales as the square-root of the typical radius of trail curvature with a strategy-dependent prefactor. By posing the problem in the reinforcement learning framework, we obtain optimal sampling strategies under various movement constraints and sensor configurations. An exactly solvable model in the Hamilton-Jacobi-Bellman framework recapitulates features of sampling strategies obtained via learning and quantifies the trade-off between movement cost, speed and sampling efficiency. Our work provides a general framework for trail tracking and testable hypotheses on the algorithms that animals use to follow trails.   

Specialization of control strategies in terrestrial slithering snakes.

Limbless locomotors like snakes use environmental heterogeneities for propulsion. We tested two snake species adapted to different habitats, the desert specialist C. occipitalis and the multi-terrain generalist P. guttatus, in a model terrestrial terrain–rigid arrays of posts on a low-friction whiteboard substrate. Principal component analysis (PCA) revealed the specialist maintained its stereotyped sand-swimming wave in the arrays, while results for the generalist were inconclusive, indicating either a periodic gait was higher dimensional, or the motion was aperiodic. Persistent homology, a mathematical technique which can identify cycles without reducing dimension, suggested the generalist used aperiodic kinematics. We hypothesized that the generalists instead controlled reaction forces and tested this using a simplified terrain, a single force-sensitive post on whiteboard. The generalist was more effective at using the post, maintaining longer contacts and more consistent force vectors. Our study suggests control specialization; the specialist targets beneficial sand swimming kinematics while the generalist controls for advantageous force generation in accord with early studies of generalist snakes in lattices [e.g. Gray 1955].   

Proceed with caution: dynamics of novelty-directed risk assessment behavior in mice

Mice encountering a novel object initially display a distinctive set of behaviors that have been described as "risk assessment," including slow extensions of their snout toward the object followed by rapid retractions. These behaviors have also been observed in mouse models of anxiety, such as during elevated plus maze tasks, and are proposed to reflect a conflict between exploration and risk avoidance. However, a simple conflict does not explain the asymmetry in the speeds of approach and retreat: why do mice spend more time approaching the object than retreating from it, when spending less time near the object overall could reduce their exposure to risk? Analyzing the snout trajectories of mice exploring a novel object, we test the hypothesis that these behaviors reflect an internal evidence accumulation process, in which mice integrate a subjective "risk" while approaching and retreat when the cumulative "risk" crosses a threshold. We ask whether 1) a feed-forward model can predict retreat timings from preceding trajectories of approach, and 2) including feedback from risk to snout velocity explains the asymmetrical dynamics of "risk assessment".   

Transitions between stochastic and oscillatory active sensing in pulse-type electric fish

Rather than wait passively for signals to be detected by their sensors, animals actively move in order to gather information from their environment. Furthermore, when sensing is performed by means of rhythmic movements, reafferent sensory streams are also rhythmic, which is advantageous for sensory processing. Here, we report on and characterize an hitherto unknown behavioural state of pulse-type weakly electric fish during which electrosensory acquisition becomes rhythmic and is coupled to low-frequency movement. The oscillatory nature of this sensory sampling strategy is in stark contrast to that exhibited during other behavioural states, which we show to be well-modelled by jump-diffusion stochastic processes.

To characterize the oscillatory dynamics of these rhythmic behavioural states, we identify and extract them from long-term observations of freely-moving fish. To achieve this, we apply a wavelet transform to the inter-pulse interval time series and use the resulting spectra as input for t-distributed Stochastic Neighbor Embedding. This creates a 2D clustered representation of the data from which we successfully classify different sensory acquisition strategies. This approach reveals that rhythmic sampling in fact occurs over a range of frequencies between 0.5 and 1 Hz.   

Steering and turning control of C. elegans

Elongate animals (e.g., snakes, nematodes) propagate waves of body curvature to generate propulsion in dissipative environments. In particular, the nematode worm C. elegans lives in environments (e.g., rotting fruit) where maneuverability is crucial to overcome heterogeneities and post-interaction deformations. To search for steering control principles in undulatory locomotion, we studied C. elegans traversing both agar and liquid buffer. These worms generate a time-dependent omega-like shape for reorientation to achieve body rotation of 150±26° on agar and 84±39° in liquid buffer. Principal component analysis (PCA) revealed the worms use four principal components during turning, superimposing two body traveling waves with two spatial frequencies. A geometric mechanics framework rationalized the observed turning dynamics by properly coupling the amplitude and the phase of the two body traveling waves. Theory predicted omega turns can achieve rotation of 153° on agar and 89° in liquid buffer, in agreement with worm experiments. These results and robophysical experiments implementing the behavior suggest that omega turns are a robust strategy for turning in diverse environments.