zoukankan      html  css  js  c++  java
  • RBM Formula Deduction

    Energy based Model

    the probability distribution (softmax function):

    [p(x)=frac{exp(-E(x))}{sumlimits_x{exp(-E(x))}}]

    when there are hidden units,

    [P(x)=sumlimits_h{P(x,h)}=frac{1}{sum_xexp(-E(x))}sumlimits_h{exp(-E(x,h))}]

     

    now, we define the free energy function:

    [F(x)=-log sumlimits_h exp(-E(x,h))]

    so that,

    [sumlimits_h exp(-E(x,h))=-exp( F(x))]

     

    now, we rewrite the probability distribution for simpilification:

    [P(x)=frac{exp(-F(x))}{sum_x{exp(-F(x))}}]

     

    then, we define the overall cost function:

    [mathcal{L}( heta,D)=-frac{1}{N}sumlimits_{x^{(i)} in D}{log p(x^{(i)})}]

    we firstly calculate the parcial gradient of $log p(x)$ with respect to $ heta$:

    [-log P(x)=F(x) + logleft(sumlimits_x{exp(-F(x))} ight)]

    [-frac{partial log P(x)}{partial heta}=frac{partial F(x)}{partial heta}-sumlimits_{hat x}{p(hat x)frac{partial F(hat x)}{partial heta}}]

    note that, the gradient contains two terms, which is called the positive phase and the negative phase. The first term increase the probability of training data, and the second term decrease the probability of samples generated by the model.

    It's difficult to determine this gradient analytically, as we can't calculate $E_P[frac{partial F(x)}{partial heta}]$. So we might estimate the expectation using sample method.

     

    we would like elements $ ilde x$ of $mathcal{N}$ to be sampled according to $P( ilde x)$, where $mathcal{N}$ is called negative particles.

    Given that, the gradient can then be written as:

    [ - frac{partial log p(x)}{partial heta}approx frac{partial F(x)}{partial heta} - frac{1}{|mathcal{N}|} sumlimits_{ ilde x in mathcal{N}}frac{partial F( ilde x)}{partial heta}]

     

    RBM

    _images/rbm.png

    the energy function $E(v,h)$ of RBM is defined as :

    [E(v,h)=-b'v-c'h-h'Wv]

    where

    • $W$ represents the weights connecting hidden and visble units.
    • $b,c$ are bias terms of visible and hidden layers respectively. 
  • 相关阅读:
    InPut 标签 HTML(表单)
    JavaScript Table 对象
    JCBD
    JCBD
    JavaScript prototype 属性
    Java8 新特性
    JavaScript 对象的使用
    Java 反射
    虚拟机VirtualBox启动虚拟机报Only Ethernet Adapter' (VERR_INTNET_FLT_IF_NOT_FOUND).
    Impala 数值函数
  • 原文地址:https://www.cnblogs.com/ZJUT-jiangnan/p/5466814.html
Copyright © 2011-2022 走看看