A lexical approach for identifying behavioural action sequences.

Area of Science:

• Ethology
• Computational Neuroscience
• Machine Learning

Background:

• Animals exhibit complex behavioral patterns during tasks, like bird spiraling or moth searching.
• Identifying rare, recurring behavioral sequences in noisy data is crucial for understanding animal responses to stimuli.
• Current models often focus on overall behavior dynamics or individual movements, not transient action sequences.

Purpose of the Study:

• To develop a novel unsupervised algorithm for identifying and segmenting rare, recurring behavioral action sequences.
• To address the limitations of existing models in detecting transient behavioral patterns in long recordings.
• To create a versatile tool applicable to diverse species and sequential data analysis.

Main Methods:

• Developed a lexical, hierarchical model of behavior.
• Designed an unsupervised algorithm named "BASS" (Behavioral Action Sequence Segmentation).
• Applied BASS to behavioral recordings of larval zebrafish and simulated glider data.

Main Results:

• BASS successfully extracted long, non-Markovian sequences from zebrafish navigation, including repeats and mixtures of forward and turn bouts.
• In a chemotaxis assay, BASS identified zebrafish strategies involving fast turns and burst swims to avoid aversive cues.
• BASS detected characteristic spiraling patterns in simulated soaring glider data, demonstrating its ability to find rare sequences.

Conclusions:

• BASS efficiently identifies and segments rare, recurring behavioral action sequences in long recordings across different contexts.
• The algorithm uncovers complex behavioral strategies, such as zebrafish chemotaxis and soaring bird patterns.
• BASS offers a broadly applicable, generic pattern recognition tool for sequential data in various scientific domains.