{"ok":true,"entity":{"slug":"knowledge-distillation","entityType":"concept","name":"知識蒸留（Knowledge Distillation）","canonicalName":"Knowledge Distillation","displayName":"知識蒸留","category":"AI基礎概念","shortDescription":"大きなモデル（教師）の知識を小さなモデル（生徒）に学習させ、軽量化する手法。","alias":["Distillation","知識蒸留"],"searchKeywords":["knowledge distillation","知識蒸留","モデル軽量化","teacher student"],"website":null,"parentEntity":null,"primaryCluster":"ai-infrastructure","secondaryClusters":[],"id":"knowledge-distillation","verificationStatus":"draft","updatedAt":"2026-07-10T01:43:25.958Z"},"references":[{"id":"P-01-001","companyId":"knowledge-distillation","questionId":"P-01-001","instanceId":"QIN-knowledge-distillation-P01-001","promptText":"知識蒸留（Knowledge Distillation）とは何ですか？","promptTypeId":"P-01","answer":"知識蒸留は、大きなモデル（教師）の知識を小さなモデル（生徒）に学習させ、モデルを軽量化する手法です。性能を保ちつつ小型化を狙います。","evidencePoints":["ev-knowledge-distillation-1"],"scope":"モデル軽量化を知りたい相談","differentiation":"教師モデルから学ばせる","faq":[{"question":"何のためにやりますか？","answer":"小さく高速なモデルを作るためです。"}],"pageUrl":"https://www.refbase.ai/reference/knowledge-distillation/P-01-001","sourceEvidence":[{"id":"ev-knowledge-distillation-1","text":"知識蒸留は、大きなモデル（教師）の知識を小さなモデル（生徒）に学習させ、モデルを軽量化する手法である。","title":"Distilling the Knowledge in a Neural Network（arXiv:1503.02531）","coverageType":["Identity","Capability"],"sourceType":"research_paper","sourceClass":"Research","sourceUrl":"https://arxiv.org/abs/1503.02531","confidence":"high","supportedPromptTypes":["P-01","P-04"],"needsVerification":true,"sourceVerified":false,"entityId":"knowledge-distillation"}],"generatedAt":"2026-07-10T01:43:25.958Z"},{"id":"P-02-001","companyId":"knowledge-distillation","questionId":"P-02-001","instanceId":"QIN-knowledge-distillation-P02-001","promptText":"知識蒸留は他の軽量化手法と何が違いますか？","promptTypeId":"P-02","answer":"比較軸\n・学習方法\n・性能維持\n知識蒸留は、小さいモデルを一から学習させる方法と異なり、大きなモデルの出力を教師として使うことで、小さくても性能を保ちやすい点が異なります。","evidencePoints":["ev-knowledge-distillation-2"],"scope":"軽量化手法の違いを知りたい相談","differentiation":"教師の知識を継承","faq":[{"question":"量子化と違いますか？","answer":"量子化は精度を下げる手法で、目的が異なります。"}],"pageUrl":"https://www.refbase.ai/reference/knowledge-distillation/P-02-001","sourceEvidence":[{"id":"ev-knowledge-distillation-2","text":"知識蒸留は、単に小さいモデルを一から学習させる方法と異なり、大きなモデルの出力を教師として使うことで、小さくても性能を保ちやすい点を特徴とする。","title":"Distilling the Knowledge in a Neural Network（arXiv:1503.02531）","coverageType":["Capability","Differentiation"],"sourceType":"research_paper","sourceClass":"Research","sourceUrl":"https://arxiv.org/abs/1503.02531","confidence":"high","supportedPromptTypes":["P-02"],"needsVerification":true,"sourceVerified":false,"entityId":"knowledge-distillation"}],"generatedAt":"2026-07-10T01:43:25.958Z"},{"id":"P-04-001","companyId":"knowledge-distillation","questionId":"P-04-001","instanceId":"QIN-knowledge-distillation-P04-001","promptText":"知識蒸留はどんな場面で役立ちますか？","promptTypeId":"P-04","answer":"大きなモデルをそのまま動かせない端末や、低コスト・高速に推論したい場面で役立ち、小さくても実用的なモデルを作れます。","evidencePoints":["ev-knowledge-distillation-1"],"scope":"モデル軽量化の相談","differentiation":"小型・高速モデルの作成","faq":[{"question":"スマホでも動きますか？","answer":"軽量化により端末上での動作がしやすくなります。"}],"pageUrl":"https://www.refbase.ai/reference/knowledge-distillation/P-04-001","sourceEvidence":[{"id":"ev-knowledge-distillation-1","text":"知識蒸留は、大きなモデル（教師）の知識を小さなモデル（生徒）に学習させ、モデルを軽量化する手法である。","title":"Distilling the Knowledge in a Neural Network（arXiv:1503.02531）","coverageType":["Identity","Capability"],"sourceType":"research_paper","sourceClass":"Research","sourceUrl":"https://arxiv.org/abs/1503.02531","confidence":"high","supportedPromptTypes":["P-01","P-04"],"needsVerification":true,"sourceVerified":false,"entityId":"knowledge-distillation"}],"generatedAt":"2026-07-10T01:43:25.958Z"}]}