Learning Beyond Full Imitation: Task-Preserving Knowledge Distillation

📄 arXiv: 2609.39338v1 📥 PDF

作者: Qianfeng Yuan, Wenbing Tao

分类: cs.LG

发布日期: 2026-09-30


💡 一句话要点

提出任务保留知识蒸馏以解决全模仿的局限性

🎯 匹配领域: 支柱二:RL算法与架构 (RL & Architecture)

关键词: 知识蒸馏 任务保留 模型压缩 深度学习 条件学习

📋 核心要点

  1. 现有知识蒸馏方法过度依赖教师模型的全模仿,导致学生模型可能丧失已有的判别能力。
  2. 论文提出任务保留知识蒸馏(TPKD),通过保持标签梯度不变并修正条件梯度来优化学生模型的学习过程。
  3. TPKD在CIFAR-100和CLINC150数据集上分别实现了88.05%和93.81%的准确率,较标准蒸馏方法有显著提升。

📝 摘要(中文)

知识蒸馏通过鼓励学生匹配教师的预测类概率来转移知识。这些概率不仅表达了对正确类的信心,还反映了错误替代类之间的关系。然而,过度模仿并不一定能提高学生的表现。本文的主要结果是明确区分全模仿与条件学习。当学生对每个错误类的概率分配不超过教师时,教师到学生的KL最小化会被阻止。重要的是,教师在错误类之间的相对概率仍然可以完全学习。我们提出了任务保留知识蒸馏(TPKD),该方法保持标签梯度不变,并最小化条件梯度的修正,以确保输出更新保留标签步骤在每个错误替代类上的增益。实验表明,TPKD在CIFAR-100和CLINC150上分别达到了88.05%和93.81%的准确率,较标准蒸馏提高了0.47和0.35个百分点。

🔬 方法详解

问题定义:本文旨在解决知识蒸馏中全模仿导致学生模型性能下降的问题。现有方法过于依赖教师模型的输出,可能使学生模型丧失自身的判别能力。

核心思路:论文提出任务保留知识蒸馏(TPKD),通过保持标签梯度不变并对条件梯度进行最小化修正,确保学生模型在学习过程中保留对正确类的判别能力。

技术框架:TPKD的整体架构包括教师模型和学生模型的交互,首先通过教师模型生成类概率,然后在学生模型中进行标签和条件匹配,最后更新学生模型的参数以保留标签增益。

关键创新:TPKD的核心创新在于明确区分全模仿与条件学习,提出在保持正确类得分优势的同时,允许学生模型学习教师在错误类之间的相对概率。这一设计使得学生模型能够在不损失判别能力的情况下进行知识转移。

关键设计:在TPKD中,损失函数设计为同时考虑标签匹配和条件匹配,确保在每一步更新中,学生模型的输出能够有效保留教师模型的知识,同时避免过度模仿带来的负面影响。

🖼️ 关键图片

img_0
img_1
img_2

📊 实验亮点

实验结果显示,TPKD在CIFAR-100数据集上达到了88.05%的准确率,在CLINC150数据集上达到了93.81%的准确率,分别较标准蒸馏方法提高了0.47和0.35个百分点,验证了其在知识转移中的有效性。

🎯 应用场景

该研究在知识蒸馏领域具有广泛的应用潜力,尤其是在需要将复杂模型的知识转移到轻量级模型的场景中,如移动设备上的实时推理和边缘计算。TPKD方法的有效性为模型压缩和加速提供了新的思路,未来可能在自动驾驶、智能家居等领域发挥重要作用。

📄 摘要(原文)

Knowledge distillation transfers knowledge by encouraging a student to match a teacher's predicted class probabilities. These probabilities express not only confidence in the correct class, but also relations among incorrect alternatives. Yet closer imitation does not necessarily yield a better student. A student may already distinguish the correct class more sharply than its teacher, so further imitation can require giving back discrimination it has acquired. Our main result is an exact separation between full imitation and conditional learning. When the correct class's score advantage over each alternative must be preserved, full teacher-to-student KL minimization is blocked exactly when the student assigns no more probability than the teacher to every incorrect class. Crucially, the teacher's relative probabilities among incorrect classes remain fully learnable. We characterize the exact price of this transfer: a minimum increase in correct-class log-odds that compensates for the largest conditional-probability mismatch. Label fitting and conditional matching can therefore be completed even as full teacher KL diverges. This separation motivates task-preserving knowledge distillation (TPKD), which keeps the label gradient intact and minimally corrects the conditional gradient so that its output update preserves the label step's gains against every incorrect alternative. The corrected conditional direction retains more than half of the original first-order conditional descent at the same step size, with a tight bound. For a fixed positive conditional target and sufficiently small constant output steps, label and conditional errors vanish together. Experiments trace this learning from exact head updates to ordinary network training. TPKD reaches 88.05% accuracy on CIFAR-100 and 93.81% on CLINC150, improving over standard distillation by 0.47 and 0.35 percentage points across three seeds.