https://feisky.xyz/machine-learning/basic.html Recent content in 机器学习概述 on Machine Learning Hugo -- gohugo.io zh-cn https://feisky.xyz/machine-learning/basic/regularization.html Mon, 01 Jan 0001 00:00:00 +0000 https://feisky.xyz/machine-learning/basic/regularization.html 在模型过于复杂的情况下，模型会学习到很多特征，从而导致可能把所有训练样本都拟合到，这样就导致了过拟合。解决过拟合可以从两个方面入手，一是减少 https://feisky.xyz/machine-learning/basic/supervised-learning.html Mon, 01 Jan 0001 00:00:00 +0000 https://feisky.xyz/machine-learning/basic/supervised-learning.html 监督学习的目标是建立一个学习过程，将预测结果与“训练数据”（即输入数据）的实际结果进行比较，不断的调整预测模型，直到模型的预测结果达到一个预 https://feisky.xyz/machine-learning/basic/load-mat.html Mon, 01 Jan 0001 00:00:00 +0000 https://feisky.xyz/machine-learning/basic/load-mat.html !mkdir -p data !wget http://ufldl.stanford.edu/housenumbers/train.tar.gz -O data/train.tar.gz !wget http://ufldl.stanford.edu/housenumbers/test.tar.gz -O data/test.tar.gz !wget http://ufldl.stanford.edu/housenumbers/extra.tar.gz -O data/extra.tar.gz !wget http://ufldl.stanford.edu/housenumbers/train_32x32.mat -O data/train_32x32.mat !wget http://ufldl.stanford.edu/housenumbers/test_32x32.mat -O data/test_32x32.mat !wget http://ufldl.stanford.edu/housenumbers/extra_32x32.mat -O data/extra_32x32.mat --2016-12-12 11:15:24-- http://ufldl.stanford.edu/housenumbers/train.tar.gz Resolving ufldl.stanford.edu (ufldl.stanford.edu)... 171.64.68.10 Connecting to ufldl.stanford.edu (ufldl.stanford.edu)|171.64.68.10|:80... connected. HTTP request sent, awaiting response... 200 OK Length: 404141560 (385M) [application/x-gzip] Saving to: 'data/train.tar.gz' 100%[======================================>] 404,141,560 6.58MB/s in 58s 2016-12-12 11:16:22 (6.62 MB/s) - 'data/train.tar.gz' saved [404141560/404141560] --2016-12-12 11:16:22-- http://ufldl.stanford.edu/housenumbers/test.tar.gz Resolving https://feisky.xyz/machine-learning/basic/datasets.html Mon, 01 Jan 0001 00:00:00 +0000 https://feisky.xyz/machine-learning/basic/datasets.html 在机器学习中，通常将所有的数据划分为三份：训练数据集、验证数据集和测试数据集。它们的功能分别为 训练数据集（train dataset）：用来构 https://feisky.xyz/machine-learning/basic/gradient-descent.html Mon, 01 Jan 0001 00:00:00 +0000 https://feisky.xyz/machine-learning/basic/gradient-descent.html 在训练机器学习模型时，首先对权重和偏差进行初始猜测，然后反复调整这些猜测，直到获得损失可能最低的权重和偏差为止（即模型收敛）。 而梯度下降是机 https://feisky.xyz/machine-learning/basic/evaluation.html Mon, 01 Jan 0001 00:00:00 +0000 https://feisky.xyz/machine-learning/basic/evaluation.html 模型在训练集上的误差通常称为 “训练误差” 或 “经验误差”，而在新样本上的误差称为 “泛化误差”。显然，机器学习的目的是得到泛化误差小的学习器。然 https://feisky.xyz/machine-learning/basic/feature-engineering.html Mon, 01 Jan 0001 00:00:00 +0000 https://feisky.xyz/machine-learning/basic/feature-engineering.html 特征工程是指从原始数据转换为特征向量的过程。特征工程是机器学习中最重要的起始步骤，会直接影响机器学习的效果，并通常需要大量的时间。典型的特征 https://feisky.xyz/machine-learning/basic/hyperparameter.html Mon, 01 Jan 0001 00:00:00 +0000 https://feisky.xyz/machine-learning/basic/hyperparameter.html 在机器学习模型中，通常训练过程会包含两种参数 模型参数，即定义模型时必需的参数，这些参数需要通过训练迭代来学习。典型的模型参数为回归模型和神经 https://feisky.xyz/machine-learning/basic/non-supervised-learning.html Mon, 01 Jan 0001 00:00:00 +0000 https://feisky.xyz/machine-learning/basic/non-supervised-learning.html 概率图模型 规则学习 聚类 集成学习