什么是 VC 维？它如何从理论上界定模型的泛化能力？
What is the VC dimension, and how does it theoretically bound a model's generalization ability?
VC 维衡量假设类能打散的最大样本数，反映模型容量。VC 理论给出泛化误差上界随 $\sqrt{d/n}$ 增长（d 为 VC 维、n 为样本数），说明容量越大越易过拟合，为模型选择与样本量需求提供了严格的理论依据。
The VC dimension measures the largest sample a hypothesis class can shatter, reflecting model capacity. VC theory gives a generalization error bound growing like $\sqrt{d/n}$ (d the VC dimension, n the sample size), showing larger capacity risks overfitting and providing a rigorous basis for model selection and sample requirements.
难度 专业 · 适合 高中(15-18)
CC BY 4.0 · 已核验 · Verified 2026-11-02
来源 · Source
《统计学习理论》Vapnik
Statistical Learning Theory by Vladimir Vapnik
独立核验依据 · Independent verification: 核验依据 1 ISBN 9787121258756（豆瓣书目条目: 统计学习理论）
#VC维 #泛化 #模型容量
什么是贝叶斯推断？它如何将先验与似然结合得到后验？
What is Bayesian inference, and how does it combine prior and likelihood into a posterior?
贝叶斯推断用贝叶斯公式 $P(\theta|D)=\frac{P(D|\theta)P(\theta)}{P(D)}$ 把先验分布 $P(\theta)$ 与观测似然 $P(D|\theta)$ 结合为后验 $P(\theta|D)$。它把参数视为随机变量，随数据更新信念，广泛应用于机器学习的不确定性估计与因果建模。
Bayesian inference uses Bayes' rule $P(\theta|D)=\frac{P(D|\theta)P(\theta)}{P(D)}$ to combine a prior $P(\theta)$ with the observed likelihood $P(D|\theta)$ into the posterior $P(\theta|D)$. Treating parameters as random variables, it updates belief with data and is widely used for uncertainty estimation and causal modeling in machine learning.
难度 专业 · 适合 高中(15-18)
CC BY 4.0 · 已核验 · Verified 2026-11-02
来源 · Source
《模式识别与机器学习》Bishop
Pattern Recognition and Machine Learning by Christopher Bishop
独立核验依据 · Independent verification: 核验依据 1 ISBN 9787115681409（豆瓣书目条目: 模式识别与机器学习）
#贝叶斯 #后验 #先验
什么是核方法？核技巧如何在不显式计算高维映射的情况下实现非线性分类？
What are kernel methods, and how does the kernel trick enable nonlinear classification without explicitly computing high-dimensional maps?
核方法把输入通过映射 $\phi$ 送入高维特征空间再线性分类。核技巧直接用核函数 $k(x,x')=\langle\phi(x),\phi(x')\rangle$ 计算内积，无需写出 $\phi$ 本身，从而以低成本获得非线性决策边界，代表算法为支持向量机（SVM）。
Kernel methods map inputs via $\phi$ into a high-dimensional feature space and classify linearly there. The kernel trick computes the inner product directly via $k(x,x')=\langle\phi(x),\phi(x')\rangle$ without ever forming $\phi$, cheaply yielding nonlinear decision boundaries; the canonical algorithm is the support vector machine (SVM).
难度 专业 · 适合 高中(15-18)
CC BY 4.0 · 已核验 · Verified 2026-10-04 · 2 条独立证据
来源 · Source
《统计学习基础》Hastie、Tibshirani 与 Friedman
The Elements of Statistical Learning by Hastie Tibshirani and Friedman
独立核验依据 · Independent verification: 核验依据 1 ISBN 9780387848846 (2009) · 核验依据 2 ISBN 9780665590955 (2002)
#核方法 #核技巧 #SVM
什么是梯度下降？它在训练神经网络中扮演什么角色？
What is gradient descent, and what role does it play in training neural networks?
梯度下降沿损失函数负梯度方向迭代更新参数 $w\leftarrow w-\eta\nabla L$，以逐步降低训练误差。它是训练神经网络的核心优化算法，配合反向传播计算梯度，学习率 $\eta$ 控制步长，是深度学习的引擎。
Gradient descent iteratively updates parameters along the negative gradient $w\leftarrow w-\eta\nabla L$ to reduce training error. It is the core optimization algorithm for training neural networks, combined with backpropagation to compute gradients, where the learning rate $\eta$ controls step size and serves as the engine of deep learning.
难度 进阶 · 适合 大众
CC BY 4.0 · 已核验 · Verified 2026-10-04
来源 · Source
《深度学习》Goodfellow、Bengio 与 Courville
Deep Learning by Goodfellow Bengio and Courville
独立核验依据 · Independent verification: 核验依据 1 ISBN 978-0-262-03561-3 (MIT Press)
#梯度下降 #反向传播 #神经网络
什么是损失函数？均方误差与交叉熵各适用于什么任务？
What is a loss function, and what tasks do mean squared error and cross-entropy suit?
损失函数量化模型预测与真实值的差距，指导优化。均方误差 $\frac1n\sum(y-\hat y)^2$ 适用于回归；交叉熵 $-\sum y\log\hat y$ 适用于分类（尤其多类概率输出），因它最大化正确类的对数似然而对错误预测惩罚重。
A loss function quantifies the gap between predictions and true values, guiding optimization. Mean squared error $\frac1n\sum(y-\hat y)^2$ suits regression; cross-entropy $-\sum y\log\hat y$ suits classification (especially multi-class probabilistic output) because it maximizes the log-likelihood of the correct class and heavily penalizes wrong predictions.
难度 进阶 · 适合 初中(12-15)
CC BY 4.0 · 已核验 · Verified 2026-10-04
来源 · Source
《深度学习》Goodfellow、Bengio 与 Courville
Deep Learning by Goodfellow Bengio and Courville
独立核验依据 · Independent verification: 核验依据 1 ISBN 978-0-262-03561-3 (MIT Press)
#损失函数 #均方误差 #交叉熵
什么是过拟合？正则化如何从数学上抑制它？
What is overfitting, and how does regularization mathematically suppress it?
过拟合是模型在训练集上误差极小却在未见数据上表现差，源于过度拟合噪声。正则化在损失函数中加惩罚项如 $\lambda\|w\|^2$（岭回归），约束参数规模，等价于在参数空间施加先验，从而降低模型复杂度、提升泛化。
Overfitting occurs when a model achieves tiny training error but poor performance on unseen data, having fit noise. Regularization adds a penalty such as $\lambda\|w\|^2$ (ridge regression) to the loss, constraining parameter magnitude and effectively imposing a prior, which reduces model complexity and improves generalization.
难度 进阶 · 适合 初中(12-15)
CC BY 4.0 · 已核验 · Verified 2026-10-04
来源 · Source
《深度学习》Goodfellow、Bengio 与 Courville
Deep Learning by Goodfellow Bengio and Courville
独立核验依据 · Independent verification: 核验依据 1 ISBN 978-0-262-03561-3 (MIT Press)
#过拟合 #正则化 #泛化