curiositech/knowledge-distillation-deep

SKILL: Knowledge Distillation and Intelligent Compression

Version: 1.0
Domain: AI/ML System Design, Agent Architecture, Knowledge Transfer
Cognitive Load: Medium (requires understanding of model training, system tradeoffs)

DECISION POINTS

Primary Decision Tree: Choosing Distillation Strategy

Given: [Task Type, Model Size Constraint, Accuracy Target]

IF task_similarity(student, teacher) > 0.8:
    ├── Use Response-Based Distillation
    │   ├── Temperature T = 3-6 for classification
    │   ├── Temperature T = 1-3 for regression
    │   └── Focus on soft label matching
    │
ELIF abstraction_level_needed == "high":
    ├── Use Feature-Based Distillation
    │   ├── Match intermediate layer activations
    │   ├── Use attention transfer if available
    │   └── Preserve semantic representations
    │
ELIF compression_ratio > 10x:
    ├── Use Cascade Architecture
    │   ├── Small model handles confidence > 0.9 cases
    │   ├── Large model backup for confidence < 0.9
    │   └── Route based on input complexity score
    │
ELSE:
    └── Use Relation-Based Distillation
        ├── Match similarity matrices between samples
        ├── Preserve ranking relationships
        └── Focus on structural knowledge transfer

Secondary Decision: Setting Alpha (Size vs Accuracy Priority)

IF deployment_environment in ["mobile", "edge", "embedded"]:
    ├── SET α = 0.6-0.8 (prioritize size)
    └── Accept accuracy loss for resource constraints

ELIF application_domain in ["safety_critical", "medical", "financial"]:
    ├── SET α = 0.1-0.3 (prioritize accuracy)
    └── Use larger models, cascade for uncertainty

ELIF can_dynamically_route == True:
    ├── SET α = 0.5 for base model
    ├── Build multiple checkpoints at different α values
    └── Route based on input difficulty estimation

ELSE:
    └── SET α = 0.4-0.6 (balanced approach)

Tertiary Decision: Agent Coordination Pattern

IF environment_stability == "static" AND task_distribution_known == True:
    └── Use Hierarchical Distillation (teacher → student)

ELIF multiple_experts_available == True:
    ├── Use Ensemble Distillation
    ├── Weight teachers by domain expertise
    └── Student learns from weighted combination

ELIF need_continuous_adaptation == True:
    ├── Use Online Distillation
    ├── Agents co-evolve simultaneously
    └── Bidirectional knowledge flow

ELSE:
    └── Use Self-Distillation with iterative refinement

FAILURE MODES

1. Accuracy Mirage

Symptoms: High average accuracy (>90%) but catastrophic failures on edge cases Diagnosis: Standard metrics hide stratified performance degradation Detection Rule: If per-class accuracy variance > 20% or minority class accuracy < 70% of overall accuracy Fix: Implement stratified validation with per-class thresholds; use weighted loss functions for rare classes

2. Temperature Blindness

Symptoms: Student model overconfident, poor calibration, loses "dark knowledge" Diagnosis: Training with temperature=1.0, ignoring probability distribution structure Detection Rule: If student confidence > teacher confidence on incorrect predictions Fix: Use temperature T=3-20 during distillation training; add calibration validation step

3. Capacity Cliff Crash

Symptoms: Model works fine until slight compression increase causes dramatic accuracy drop Diagnosis: Hit minimum capacity threshold for task complexity Detection Rule: If >5% accuracy drop from <10% parameter reduction Fix: Set hard minimum model size; use pruning instead of architecture changes; implement cascade routing

4. Hierarchical Rigidity

Symptoms: System can't adapt to new patterns; student errors persist despite available corrections Diagnosis: Fixed teacher-student roles prevent bidirectional learning Detection Rule: If student discovers edge cases but can't update teacher knowledge Fix: Implement feedback loops; use ensemble loss functions; enable peer learning between agents

5. Compression Amplification Bias

Symptoms: Compressed model maintains average performance but amplifies existing biases Diagnosis: Distillation preserves teacher's biases while losing error correction capacity Detection Rule: If demographic parity decreases >10% or fairness metrics degrade disproportionately Fix: Use bias-aware distillation loss; oversample minority classes; validate on adversarial fairness benchmarks

WORKED EXAMPLES

Example 1: Mobile Deployment Scenario

Context: Deploying sentiment analysis to mobile app, 50MB model limit, 200ms latency requirement Teacher: 800MB BERT model, 95% accuracy, 1.2s inference Constraints: α = 0.7 (heavily prioritize size), maintain >90% accuracy

Decision Process:

1. Task similarity high (same domain) → Response-based distillation
2. Compression ratio 16x → Expect significant accuracy loss, need mitigation
3. Temperature selection: T=4 for sentiment (discrete classes with similarity structure)

Implementation:

• Train 50MB DistilBERT student on soft labels from BERT teacher
• Temperature T=4 during training, T=1 during inference
• Achieve 91.2% accuracy (3.8% drop) with 16x compression
• Add uncertainty threshold: defer to cloud API when confidence <0.85

Outcome: DS = 0.7×(50/800) + 0.3×(1-91.2/95) = 0.7×0.0625 + 0.3×0.04 = 0.056 (excellent score) Trade-off: 3.8% accuracy loss for 16x size reduction and 10x speed improvement

Example 2: Safety-Critical Multi-Agent System

Context: Autonomous vehicle perception, multiple specialized agents for object detection, depth estimation, trajectory planning Constraints: α = 0.2 (heavily prioritize accuracy), 99.9% reliability requirement

Decision Process:

1. Different tasks per agent → Feature-based distillation for shared representations
2. Safety critical → Cascade architecture with redundancy
3. Multi-expert system → Ensemble distillation with specialist weighting

Implementation:

• Large teacher models for each perception task
• Medium student agents learn shared feature representations
• Tiny monitoring agent validates cross-agent consistency
• Cascade: students handle confidence >0.95, teachers handle edge cases

Agent Architecture:

• Object detection student: 100MB (from 500MB teacher)
• Depth estimation student: 80MB (from 400MB teacher)
• Trajectory planning: Keep full teacher (no compression for final decisions)
• Monitor agent: 10MB, watches for inconsistencies

Outcome: System maintains 99.92% safety threshold while reducing compute by 60% Trade-off: Modest efficiency gain for maintained safety with reduced single points of failure

Example 3: Collaborative Research Assistant Agents

Context: Multi-agent system for scientific literature analysis, agents specialize in different domains but share knowledge Constraints: Dynamic environment, new papers daily, agents must learn from each other

Decision Process:

1. Dynamic environment → Online distillation with peer learning
2. Different specializations → Relation-based distillation for structural knowledge
3. Continuous adaptation needed → Bidirectional knowledge flow

Implementation:

• 5 specialist agents (biology, chemistry, physics, computer science, medicine)
• Each agent maintains domain expertise but learns general patterns from peers
• Weekly ensemble sessions where agents teach each other via attention transfer
• Self-distillation within each agent to compress learned knowledge

Mechanism:

• Agent A discovers new pattern in biology papers
• Shares attention weights and feature representations with other agents
• Other agents evaluate if pattern applies to their domains
• Successful transfers update shared knowledge base

Outcome: Collective accuracy improves 8% over 6 months vs. isolated training Trade-off: Increased coordination complexity for better adaptation and knowledge sharing

QUALITY GATES

• [ ] Distillation Score (DS) calculated with explicit α parameter and documented rationale
• [ ] Stratified validation completed with per-class accuracy thresholds set and met
• [ ] Temperature parameter tuned (T>1 during training) and calibration validated
• [ ] Capacity cliff analysis performed - confirmed model size is above minimum threshold
• [ ] Failure mode monitoring implemented for bias amplification and edge case degradation
• [ ] Adversarial test suite created covering out-of-distribution and minority class scenarios
• [ ] If compression ratio >5x, cascade architecture evaluated and routing strategy defined
• [ ] If multi-agent system, coordination pattern chosen and feedback loops implemented
• [ ] Deployment readiness confirmed with latency, memory, and accuracy benchmarks met
• [ ] Rollback plan prepared with performance monitoring alerts and degradation thresholds

NOT-FOR BOUNDARIES

Do NOT use this skill for:

• Simple rule-based systems with <1000 parameters (use direct optimization instead)
• Single-task agents with abundant computational resources (compression unnecessary)
• Systems where interpretability is primary requirement (use interpretable-models skill instead)
• Real-time systems with <10ms latency requirements (use hardware-acceleration skill instead)
• Prototype/research phases before performance requirements defined (premature optimization)

Delegate to other skills:

• For model architecture selection → use neural-architecture-search skill
• For hardware optimization → use model-deployment-optimization skill
• For interpretable AI requirements → use explainable-ai-design skill
• For federated learning scenarios → use distributed-learning-coordination skill
• For adversarial robustness → use adversarial-defense-strategies skill