DA就是“Denoising Autoencoders”的缩写。继续给yusugomori做注释，边注释边学习。看了一些DA的材料，基本上都在前面“转载”了。学习中间总有个疑问：DA和RBM到底啥区别？（别笑，我不是“学院派”的看Deep Learning理论，如果“顺次”看下来，可能不会有这个问题），现在了解的差不多了，详情见：【deep learning学习笔记】Autoencoder。之后，又有个疑问，DA具体的权重更新公式是怎么推导出来的？我知道是BP算法，不过具体公示的推导、偏导数的求解，没有看到哪个材料有具体的公式，所以姑且认为yusugomori的代码写的是正确的。
注释后的头文件：
// The Class of denoising auto-encoder
class dA
{
public:
int N; // the number of training samples
int n_visible; // the number of visible nodes
int n_hidden; // the number of hidden nodes
double **W; // the weight connecting visible node and hidden node
double *hbias; // the bias of hidden nodes
double *vbias; // the bias of visible nodes public:
// initialize the parameters
dA ( int, // N
int, // n_visible
int , // n_hidden
double**, // W
double*, // hbias
double* // vbias
);
~dA(); // make the input noised
void get_corrupted_input (
int*, // the original input 0-1 vector -- input
int*, // the resulted 0-1 vector gotten noised -- output
double // the p probability of noise, binomial test -- input
);
// encode process: calculate the probability output from hidden node
// p(hi|v) = sigmod ( sum_j(vj * wij) + bi), it's same with RBM
// but different from RBM, it dose not generate 0-1 state from Bernoulli distribution
void get_hidden_values (
int*, // the input from visible nodes
double* // the output of hidden nodes
);
// decode process: calculate the probability output from visiable node
// p(vi|h) = sigmod ( sum_j(hj * wij) + ci), it's same with RBM
// but different from RBM, it dose not generate 0-1 state from Bernoulli distribution
void get_reconstructed_input (
double*, // the input from hidden nodes
double* // the output reconstructed of visible nodes
);
// train the model by a single sample
void train (
int*, // the input sample from visiable node
double, // the learning rate
double // corruption_level is the probability of noise
);
// reconstruct the input sample
void reconstruct (
int*, // the input sample -- input
double* // the reconstructed value -- output
);
};