UFLDL Tutorial - Supervised Learning and Optimization

UFLDL Tutorial

原始代码可以从这里（GitHub repository）一次性下载。需要注意的是有些数据需要自己去下载，比如，在做PCA的练习时，需要下载MNIST数据集，可以到THE MNIST DATABASE下载。

---

Supervised Learning and Optimization

Linear Regression

Exercise 1A：线性回归预测房价，只需补充目标函数及其梯度，计算公式见原网页 ：

补充代码

linear_regression.m

%%% YOUR CODE HERE %%%
theta = theta';
%Compute the linear regression objective
for j = 1:m
    f = f + (theta*X(:,j) - y(j))^2;
end
f = f/2;

%Compute the gradient of the objective
for j = 1:m
    g = g + X(:,j)*(theta*X(:,j) - y(j));
end

实验结果

如原文所述，训练和测试误差一般在$4.5$和$5$之间，本人实验结果：

Optimization took 1.780584 seconds.
RMS training error: 4.731236
RMS testing error: 4.584099

Logistic Regression

上面的线性回归有两个特点：

• 预测连续值（房价）；
• 输出是输入的 线性函数（$y=h_{\theta }(x)={\theta }^{T}x$）；
• 代价函数为均方误差函数。

Logistic Regression：

• 预测离散值，通常用于分类；
• 输出是输入的非线性函数（sigmoid或Logistic 函数：$y=h_{\theta }(x)=\sigma ({\theta }^{T}x)$，$\sigma (z)=\frac{1}{1+exp(-z)}$）；
• 代价函数取交叉熵（概率模型推导，最大似然，见CS229 Notes）。

Exercise 1B：Logistic 分类，用于手写体。只需补充目标函数及其梯度，计算公式见原网页，推导见CS229 Notes：

补充代码
与线性回归基本相同，只是假设$y=h_{\theta }(x)=\sigma ({\theta }^{T}x)$，$\sigma (z)=\frac{1}{1+exp(-z)}$为sigmoid函数，不再是线性函数。

%%% YOUR CODE HERE %%%

%Compute the linear regression objective and it's gradient
for j = 1:m
    coItem = sigmoid(theta'*X(:,j));
    f = f - y(j)*log(coItem) - (1-y(j))*log(1-coItem);
    g = g + X(:,j)*(coItem-y(j));
end

实验结果

如原网页所述，最终训练和测试精度都为100%，本人实验结果：

Optimization took 15.115248 seconds.
Training accuracy: 100.0%
Test accuracy: 100.0%

Vectorization

补充代码

需要取消ex1a_linreg.m和ex1b_logreg.m文件中下面的注释：
ex1a_linreg.m

% theta = rand(n,1);
% tic;
% theta = minFunc(@linear_regression_vec, theta, options, train.X, train.y);
% fprintf('Optimization took %f seconds.\n', toc);

ex1b_logreg.m

% theta = rand(n,1)*0.001;
% tic;
% theta=minFunc(@logistic_regression_vec, theta, options, train.X, train.y);
% fprintf('Optimization took %f seconds.\n', toc);

linear_regression_vec.m

%%% YOUR CODE HERE %%%
f = (norm(theta'*X - y))^2 / 2;
g = X*(theta'*X-y)';

logistic_regression_vec.m

%%% YOUR CODE HERE %%%
coItem = sigmoid(theta'*X);
f = -log(coItem)*y' -log(1-coItem)*(1-y)';
g = X*(coItem-y)';

实验结果

速度快了好些，如下：

线性回归：

Optimization took 0.032485 seconds.(矢量化前约0.3s)
RMS training error: 4.023758
RMS testing error: 6.783703

Logistic分类：

Optimization took 3.419164 seconds.（矢量化前约12s）
Training accuracy: 100.0%
Test accuracy: 100.0%

Debugging: Gradient Checking

补充代码
下面是一次进行上述线性回归Logistic分类练习的梯度检验代码
grad_check_demo.m

%% for linear regression

% Load housing data from file.
data = load('housing.data');
data=data'; % put examples in columns

% Include a row of 1s as an additional intercept feature.
data = [ ones(1,size(data,2)); data ];

% Shuffle examples.
data = data(:, randperm(size(data,2)));

% Split into train and test sets
% The last row of 'data' is the median home price.
train.X = data(1:end-1,1:400);
train.y = data(end,1:400);

test.X = data(1:end-1,401:end);
test.y = data(end,401:end);

m=size(train.X,2);
n=size(train.X,1);

% Initialize the coefficient vector theta to random values.
theta0 = rand(n,1);

num_checks = 20;
% without vectorize
average_error = grad_check(@linear_regression, theta0, num_checks, train.X, train.y)
% vectorize
average_error = grad_check(@linear_regression_vec, theta0, num_checks, train.X, train.y)

%% for Logistic Classification
binary_digits = true;
[train,test] = ex1_load_mnist(binary_digits);

% Add row of 1s to the dataset to act as an intercept term.
train.X = [ones(1,size(train.X,2)); train.X]; 
test.X = [ones(1,size(test.X,2)); test.X];

% Training set dimensions
m=size(train.X,2);
n=size(train.X,1);

% Train logistic regression classifier using minFunc
options = struct('MaxIter', 100);

% First, we initialize theta to some small random values.
theta0 = rand(n,1)*0.001;

num_checks = 20;
% without vectorize
average_error = grad_check(@logistic_regression, theta0, num_checks, train.X, train.y)
% vectorize
average_error = grad_check(@logistic_regression_vec, theta0, num_checks, train.X, train.y)

实验结果

验证20次的平均误差分别为：

1.7030e-05（linear）
1.2627e-05（linear_vec）
6.0687e-06（Logistic）
8.1527e-06（Logistic_vec）

Softmax Regression

多分类，Logistic回归的推广。

---