| name | knowledge-distillation-in-deep-learning |
| description | Design deployment-focused distillation systems that balance model size, accuracy, calibration, and cascade escalation under real resource limits. Best for teacher-student compression, threshold design, and failure-aware deployment. Activate on "model compression", "teacher- student", "distillation score", "cascade model", "edge deployment", or "model calibration". NOT for generic deep-learning overviews, prompt optimization, or training work without a concrete distillation objective. |
| license | Apache-2.0 |
| allowed-tools | ["Read","Write","Edit","Glob","Grep"] |
| metadata | {"category":"Research & Academic","tags":["knowledge-distillation","model-compression","deployment","calibration","cascades","deep-learning"],"pairs-with":[{"skill":"knowledge-distillation-a-survey","reason":"Use the survey companion when the open question is transfer strategy rather than deployment policy."},{"skill":"llm-router","reason":"Distillation cascades and escalation thresholds map naturally onto runtime routing design."}],"provenance":{"kind":"legacy-recovered","sourceDocument":"Knowledge Distillation in Deep Learning and Its Applications","sourceAuthors":["Abdolmaged Alkhulaifi","Fahad Alsahli","Irfan Ahmad"],"sourceArtifact":".claude/skills/knowledge-distillation-in-deep-learning/_book_identity.json","importedFrom":"legacy-recovery","owners":["some-claude-skills"]},"authorship":{"authors":["Abdolmaged Alkhulaifi","Fahad Alsahli","Irfan Ahmad"],"maintainers":["some-claude-skills"]}} |
Knowledge Distillation In Deep Learning
Use this skill when the core design question is how to compress a capable model into something smaller, cheaper, or more routable without hiding unacceptable losses in calibration, robustness, or edge-case performance.
When to Use
- You are designing a teacher-student compression pipeline for deployment-constrained inference.
- You need to set explicit size versus accuracy priorities instead of hand-waving about "good enough."
- A small model should handle common cases first, with escalation to a larger backup model when uncertainty rises.
- You need to reason about whether compressed performance is failing because of capacity cliffs, bias amplification, or poor calibration.
- The system architecture needs different-capacity agents for different abstraction levels.
NOT for Boundaries
This skill is not the primary fit for:
- Generic deep-learning overviews or literature surveys with no concrete distillation objective.
- Prompt optimization, retrieval tuning, or non-neural transfer problems.
- Compression discussions that stop at average accuracy and ignore calibration, minority slices, or out-of-distribution behavior.
- Architecture work where there is no teacher-student relationship, cascade, or compression tradeoff to manage.
Core Mental Models
Knowledge Is Multi-Representational
Distillation can transfer logits, soft labels, feature maps, or other internal structure. Deployment success depends on choosing the representation that preserves the behavior you actually care about.
Size-Accuracy Tradeoffs Need An Explicit Policy
The paper's distillation-score framing matters because it forces the team to say what it is optimizing for. Mobile and edge deployments tolerate more loss than safety-critical paths. If the priority is hidden, the system will drift toward accidental risk.
Abstraction Level Should Drive Capacity
Low-level feature extraction can often live in smaller students. High-level semantic or reasoning-heavy decisions usually require more capacity. Uniform compression across all layers or roles is usually the wrong move.
Cascades Beat Wishful Compression
When the input distribution has easy and hard cases, a small-first, large-backup cascade often dominates "one tiny model for everything." Distillation should support escalation, not eliminate it.
Compression Amplifies Existing Weaknesses
Compressed models often fail first on rare classes, borderline examples, and out-of-distribution inputs. Average accuracy can stay high while the risk profile becomes much worse.
Decision Points
See the richer flow, quadrant, and sequence diagrams in .
