Custom behavior labels per frame
Your user-defined behavior categories applied frame-by-frame with confidence scores. Same CSV format as built-in classifiers.
CSV
← Capabilities
Custom Model Training
Train classifiers on your own behavioral data
When pretrained models do not capture your lab-specific behaviors, train custom classifiers on your own annotated sessions and validate against held-out data.
Custom
User-defined behavior labels
Your data
Training pipeline
YOLO
Detection model fine-tuning
Export
Trained models portable
The problem
Pretrained models miss lab-specific behaviors
Every lab has unique behavioral categories — specific grooming subtypes, custom task-related behaviors, or species-specific actions that general-purpose classifiers do not recognize.
- General classifiers lack categories for novel or rare behaviors
- Adapting to new species requires retraining from scratch
- No pathway from manual annotation to automated classification
The solution
Your annotations become automated classifiers
Annotate behavior in a subset of sessions, train a custom classifier, and validate on held-out data. Once accuracy is verified, apply the classifier to your full dataset.
- Annotation interface built into the ConductVision workflow
- Train on as few as 3-5 annotated sessions for common behaviors
- Validation metrics reported before deployment to full dataset
Endpoints
Custom classifiers produce the same structured output as built-in models.
Training performance report
Accuracy, precision, and recall per behavior class on held-out validation data. Confusion matrix included.
CSVPDF
Exportable trained model
Trained model weights saved for reuse across sessions, computers, and collaborators. Version-controlled with training metadata.
PyTorch model file
Applications
Custom training extends ConductVision to any behavioral vocabulary.
Lab-specific ethogram automation
Define your own behavioral categories — tail rattling, specific grooming subtypes, task-specific actions — and train classifiers that score them automatically.
Measures
- Custom labels
- Per-class accuracy
- Bout analysis
Detect low-frequency events across large datasets
Train on annotated examples of rare behaviors (seizures, stereotypies, specific social events) and scan entire datasets automatically.
Measures
- Detection sensitivity
- False positive rate
- Event count
Extend tracking to non-standard species
Fine-tune detection and pose models for species not covered by pretrained models — custom arthropods, fish species, or non-standard rodent strains.
Measures
- Species-specific accuracy
- Keypoint detection
- Model transfer
Share trained models across collaborating labs
Export trained classifiers and share with collaborators to ensure consistent scoring across sites in multi-center studies.
Measures
- Cross-site consistency
- Model portability
- Version control
Related capabilities
Behavior Classification
ML-powered frame-level behavioral classification with temporal smoothing and custom classifier training.
Model Benchmarking
Per-paradigm accuracy benchmarks with published precision, recall, and mAP scores across all 24 supported tests.
Unsupervised Discovery
Unsupervised behavioral clustering reveals structure without predefined ethograms — UMAP visualization and novel behavior flagging.
Train a classifier on your own behavioral data
Download the trial, annotate a few sessions, and build a custom classifier for your specific ethogram.
