Roadmap

🔜 Q2 2025

• Add support for SLEAP (#8)
• Add support for saving and loading HMMs (#14)
• Speed up HMM fitting by training on a random subset of data (#22)
• Add native support for parallel HMM scanning in CLI (#23)
• Test training LISBET classifiers using HMM prototypes (#24)
• Add native support for n > 2 individuals (#28)
• Enable segmentation mask generation from keypoints (#38
• Testing “multi-dyadic” strategy for inference on groups (n > 2) (#27)

📅 Q3 2025

• Introduce “Switch Individuals” as a data augmentation strategy
• Conduct ablation study on pretraining tasks

🔮 Future Ideas

• Explore end-to-end models for discovery-driven behavior labeling (#21)
• Add support for raw video input
• Add support for segmentation mask input
• Compare keypoints vs. keypoint-derived segmentation mask performance
• Evaluate frame-based models (raw video and segmentation masks) for n > 2 individuals
• Refactor CLI (#20)
• Introduce GUI
• Explore the use of LLMs for behavior classification from embeddings
• Automate model benchmarking