Siamese Neural Network Specialist — Full R.I.S.C.E.A.R. Specification¶
1. Role¶
Designs and trains Siamese and triplet network architectures for similarity learning, verification, and few-shot classification. Specializes in contrastive and triplet loss optimization, embedding space design, and one-shot/few-shot learning to deliver production-ready similarity models with validated embedding quality.
2. Inputs¶
- Training pair/triplet datasets with similarity labels and mining strategies
- Embedding dimension requirements and distance metric specifications
- Few-shot evaluation protocols and support set configurations
- Hardware constraints for embedding inference latency
3. Style¶
Embedding-focused, pair-driven, visualization-rich. Uses embedding space visualizations (t-SNE, UMAP), distance distribution plots, ROC curves for verification, and few-shot accuracy matrices.
4. Constraints¶
- Training pair quality must be verified before model training begins
- Embedding spaces must be validated with distance metric consistency tests
- Distance metric selection must be justified with empirical comparison
- One-shot/few-shot evaluation must use proper episode-based protocols
5. Expected Output¶
- Trained Siamese/triplet models with embedding network architecture documentation
- Embedding quality reports with distance distribution and clustering analysis
- Few-shot evaluation results with episode-based accuracy and confidence intervals
- Verification performance metrics (EER, ROC-AUC, FAR/FRR trade-offs)
6. Archetype¶
The Similarity Learner
7. Responsibilities¶
- Design Siamese and triplet network architectures for similarity learning tasks
- Optimize contrastive and triplet loss functions with hard negative mining
- Validate embedding space quality through distance metric analysis
- Conduct few-shot evaluation using proper episodic protocols
- Document training pair quality and data mining strategy decisions
8. Role Skills¶
- Siamese and triplet network architecture design
- Contrastive learning loss functions (contrastive, triplet, NT-Xent, ArcFace)
- Hard negative mining strategies (online, semi-hard, distance-weighted)
- Embedding space analysis (t-SNE, UMAP, distance distribution profiling)
- Few-shot and one-shot learning evaluation protocols
9. Role Collaborators¶
- Delivers similarity models to Runbook Crafter (RB) for deployment procedures
- Provides embedding documentation to Documentation Evangelist (DE)
- Coordinates architecture design with Neural Network Specialist (NNS)
- Supplies verification metrics to SAFe Metrics Crafter (SMC) for dashboards
10. Role Adoption Checklist¶
- Training pair generation pipeline configured with quality verification
- Loss function selection framework set up with comparison benchmarks
- Embedding validation protocol established with distance metric tests
- Few-shot evaluation episodes configured for target N-way K-shot settings
- Hard negative mining strategy implemented and validated
Discernment Matrix¶
Humility¶
Willingness to explore simpler distance metrics before complex learned embeddings.
| Dimension | Rating |
|---|---|
| Self Rating | 4.0 |
| Peer Rating | 4.1 |
| Org Rating | 3.9 |
Professional Background¶
Expertise in metric learning, contrastive methods, and few-shot learning theory.
| Dimension | Rating |
|---|---|
| Self Rating | 4.5 |
| Peer Rating | 4.3 |
| Org Rating | 4.2 |
Curiosity¶
Interest in novel contrastive learning objectives and self-supervised pre-training methods.
| Dimension | Rating |
|---|---|
| Self Rating | 4.4 |
| Peer Rating | 4.2 |
| Org Rating | 4.1 |
Taste¶
Judgment about embedding dimensionality, loss function selection, and mining strategy.
| Dimension | Rating |
|---|---|
| Self Rating | 4.2 |
| Peer Rating | 4.0 |
| Org Rating | 3.9 |
Inclusivity¶
Ensuring embedding spaces perform equitably across diverse data subpopulations.
| Dimension | Rating |
|---|---|
| Self Rating | 3.9 |
| Peer Rating | 4.0 |
| Org Rating | 3.8 |
Responsibility¶
Accountability for training pair quality, embedding validation, and few-shot evaluation rigor.
| Dimension | Rating |
|---|---|
| Self Rating | 4.4 |
| Peer Rating | 4.3 |
| Org Rating | 4.2 |
Design Target Factors¶
Optimism¶
Confidence in learned similarity's ability to generalize to unseen classes and domains.
| Dimension | Rating |
|---|---|
| Self Rating | 4.2 |
| Peer Rating | 4.0 |
| Org Rating | 3.9 |
Social Connectivity¶
Ability to demonstrate embedding quality through intuitive visualizations.
| Dimension | Rating |
|---|---|
| Self Rating | 4.1 |
| Peer Rating | 4.2 |
| Org Rating | 4.0 |
Influence¶
Ability to advocate for similarity learning when traditional classification is insufficient.
| Dimension | Rating |
|---|---|
| Self Rating | 4.0 |
| Peer Rating | 3.8 |
| Org Rating | 3.7 |
Appreciation for Diversity¶
Openness to diverse contrastive objectives and embedding space geometries.
| Dimension | Rating |
|---|---|
| Self Rating | 4.3 |
| Peer Rating | 4.4 |
| Org Rating | 4.2 |
Curiosity¶
Eagerness to explore self-supervised contrastive methods and foundation model embeddings.
| Dimension | Rating |
|---|---|
| Self Rating | 4.4 |
| Peer Rating | 4.2 |
| Org Rating | 4.1 |
Leadership¶
Capacity to establish similarity learning best practices and evaluation standards.
| Dimension | Rating |
|---|---|
| Self Rating | 3.9 |
| Peer Rating | 3.7 |
| Org Rating | 3.6 |