A lexical approach for identifying behavioral action sequences

Gautam Reddy, Laura Desban, Hidenori Tanaka, Julian Roussel, Olivier Mirat, Claire WyartbioRxiv 2020 · DOI

TL;DR

This paper introduces BASS, an unsupervised algorithm that identifies rare, recurring behavioral action sequences in noisy animal movement data. By treating behavior as a "language" of motifs, BASS can uncover complex strategies, like zebrafish avoiding toxins or gliders soaring, that are missed by traditional methods.

What Problem Does This Paper Address?

Understanding animal behavior often involves identifying patterns. While we can track animal movements and segment them into basic "locomotor episodes" (like a forward swim or a turn), the real meaning often lies in sequences of these episodes. These sequences can be rare, transient, and buried in noisy data, making them hard to find with existing methods. Current approaches often focus on the dynamics of individual episodes or simple transitions, failing to capture longer, more complex behavioral "words" or "phrases" that are crucial for understanding how animals respond to their environment, especially when those responses are subtle or infrequent.

Methods

The core of this work is the BASS (Behavioral Action Sequence Segmentation) algorithm. BASS is an unsupervised, hierarchical, and lexical model of behavior.

  • Lexical Model: It treats behavior as a sequence of "words" (motifs) drawn from a "dictionary." Each motif is a sequence of elementary locomotor episodes (like "forward swim, turn, forward swim").
  • Hierarchical Model: It uses a two-level approach. The lower level maps observed movements to a latent state space (like a Hidden Markov Model), and the upper level models noisy instantiations of motifs.
  • Unsupervised Learning: BASS learns the dictionary of motifs and their probabilities directly from the data without prior knowledge of stimuli or specific behaviors.
  • Generative Process: The model assumes animals generate behavior by sequentially drawing motifs from a dictionary and then "mutating" them (e.g., through insertions or deletions) before producing observable actions. This accounts for "pattern noise" (variations in motif execution) and "locomotor episode noise" (variations in individual movements).
  • Inference: BASS uses an iterative procedure to estimate motif probabilities and expand the dictionary by identifying pairs of motifs that co-occur more often than expected.

Datasets

  • Larval Zebrafish Exploration Data: Approximately 85,000 bouts from 171 fish swimming in elongated swim arenas.
  • Larval Zebrafish Chemotaxis Data: Approximately 66,000 bouts from 135 fish in a novel chemotaxis assay designed to elicit avoidance of aversive cues (acidic pH).
  • Simulated Soaring Glider Data: Simulated trajectories of a glider executing a thermalling strategy in a convective environment.
  • Previously Published Zebrafish Data: Data from Marques et al. [16] was used to validate findings on exploratory behavior.

Key Insights and Ideas

  • Behavior as a Language: The analogy of behavior as a language, with motifs as words and sequences as sentences, provides a powerful framework for uncovering complex, long-term behavioral structures.
  • Rare Events Matter: BASS's ability to identify rare but behaviorally relevant sequences is crucial for understanding subtle responses that might be drowned out by more frequent, less informative behaviors.
  • Non-Markovian Behavior: The paper demonstrates that animal behavior, particularly in complex tasks, often deviates from simple Markovian models, with longer-range dependencies and motif structures being key.
  • Dichotomy in Bout Types: In both exploration and aversive responses, zebrafish larvae tend to use sequences composed of either slow bouts (forward swims, turns) or fast bouts (bursts, fast forward swims), but rarely mix them within a single motif.
  • Stimulus-Independent Discovery: BASS can identify meaningful behavioral sequences even when the specific sensory stimuli driving them are unknown or complex.

Key Results

  • Zebrafish Exploration: BASS identified motifs that covered, on average, 78% of all bouts, revealing that a significant portion of exploratory behavior is structured into recurring sequences. Many of these motifs were long repeats of the same bout type or mixtures of slow bouts or fast bouts.
  • Zebrafish Chemotaxis: In response to an aversive acidic gradient, BASS uncovered action sequences of fast bursts and large avoidance turns that were significantly over-represented compared to exploration. These sequences made up a small fraction (~0.2%) of the dataset but were critical for avoiding the noxious environment.
  • Soaring Glider Simulation: BASS successfully identified characteristic long spiraling patterns in simulated glider trajectories, demonstrating its applicability to diverse behaviors.
  • Model Performance: BASS showed near-optimal segmentation performance on synthetic data with increasing dataset size and successfully reconstructed motif probabilities despite noise.

Limitations and Future Work

  • Computational Scalability: While BASS can handle datasets of up to 300,000 bouts and dictionaries of 500 motifs, approximations may be needed for larger datasets.
  • Continuous Motor Output: The current BASS model is designed for discrete locomotor episodes. The authors suggest future work could extend it to handle continuous motor output by adapting the elementary episode definition and using hierarchical clustering on model space.
  • Prior Knowledge Integration: While BASS is unsupervised, future work could explore incorporating prior knowledge (e.g., on motif lengths or frequencies) through Bayesian approaches (MAP estimation) or Dirichlet priors.
  • Complex Stimulus-Response: For situations where a single stimulus-response map yields highly diverse outcomes, inference might be difficult without access to the input stream, suggesting the need for new algorithmic approaches.
  • Generalization to Other Species: While BASS is presented as a generic algorithm, its application and validation across a wider range of species and behaviors would be valuable.

Summary updated

Abstract

Animals display characteristic behavioral patterns when performing a task, such as the spiraling of a soaring bird or the surge-and-cast of a male moth searching for a female. Identifying such recurring sequences occurring rarely in noisy behavioral data is key to understanding the behavioral response to a distributed stimulus in unrestrained animals. Existing models seek to describe the dynamics of behavior or segment individual locomotor episodes rather than to identify the rare and transient sequences of locomotor episodes that make up the behavioral response. To fill this gap, we develop a lexical, hierarchical model of behavior. We designed an unsupervised algorithm called “BASS” to efficiently identify and segment recurring behavioral action sequences transiently occurring in long behavioral recordings. When applied to navigating larval zebrafish, BASS extracts a dictionary of remarkably long, non-Markovian sequences consisting of repeats and mixtures of slow forward and turn bouts. Applied to a novel chemotaxis assay, BASS uncovers chemotactic strategies deployed by zebrafish to avoid aversive cues consisting of sequences of fast large-angle turns and burst swims. In a simulated dataset of soaring gliders climbing thermals, BASS finds the spiraling patterns characteristic of soaring behavior. In both cases, BASS succeeds in identifying rare action sequences in the behavior deployed by freely moving animals. BASS can be easily incorporated into the pipelines of existing behavioral analyses across diverse species, and even more broadly used as a generic algorithm for pattern recognition in low-dimensional sequential data.
BibTeX
@misc{reddy2020lexical,
  author = {Gautam Reddy and Laura Desban and Hidenori Tanaka and Julian Roussel and Olivier Mirat and Claire Wyart},
  title = {{A lexical approach for identifying behavioral action sequences}},
  year = {2020},
  doi = {10.1101/2020.08.27.270694},
  howpublished = {bioRxiv},
}

Loading related content...

Figure