1. Optimization Procedure

A good choice for the criterion is maximum likelihood regularized with dropout, possibly also with weight decay.

A good choice for the optimization algorithm for a feed-forward network is usually stochastic gradient descent with momentum.

2. Loss Function and Conditional Log-Likelihood

In the 80’s and 90’s the most commonly used loss function was the squared error

L(fθ(x),y)=fθ(x)y2 L({ f }_{ \theta }(x),y)={ ||f\theta(x)-y|| }^{ 2 }

if f is unrestricted (non-parametric),

f(x)=E[yx=x] f(x) = E[y | x = x]

Replacing the squared error by an absolute value makes the neural network try to estimate not the conditional expectation but the conditional median

分类交叉熵损失(Categorical Cross-Entropy Loss)。

交叉熵(cross entropy)目标函数 : when y is a discrete label, i.e., for classification problems, other loss functions such as the Bernoulli negative log-likelihood have been found to be more appropriate than the squared error. (y0,1y\in{ { 0,1 } })

L(fθ(x),y)=ylogfθ(x)(1y)log(1fθ(x)) L({ f }_{ \theta }(x),y)=-ylog{ f }_{ \theta }(x)-(1-y)log(1-{ f }_{ \theta }(x))

fθ(x){f}_{\theta}(x) to be strictly between 0 to 1: use the sigmoid as non-linearity for the output layer(matches well with the binomial negative log-likelihood cost function)

The mean is halved(12\frac 12)as a convenience for the computation of the gradient descent, as the derivative term of the square function will cancel out the(12\frac 12) term.

3. Learning a Conditional Probability Model

负对数似然(NLL:Negative Log Likelihood)

loss function as corresponding to a conditional log-likelihood, i.e., the negative log-likelihood (NLL) cost function

LNLL(fθ(x),y)=logP(y=yx=x;θ) {L}_{NLL}({f}_{\theta}(x),y)=-logP(y=y|x=x;\theta)

Example: if y is a continuous random variable and we assume that, given xx, it has a Gaussian distribution with mean fθ(x){f}_{\theta}(x) and variance σ2{\sigma}^{2}

logP(yx;θ)=12(fθ(x)y)1/σ2+log(2πσ2) -logP(y|x;\theta)=\frac { 1 }{ 2 } { ({ f }_{ \theta }(x)-y) }^{ 1 }/{ \sigma }^{ 2 }+log(2\pi{\sigma}^{ 2 })

Minimizing this negative log-likelihood is therefore equivalent to minimizing the squared error loss.

For discrete variables, the binomial negative log-likelihood cost function corresponds to the conditional log-likelihood associated with the Bernoulli distribution (also known as cross entropy) with probability p=fθ(x)p = {f}_{\theta}(x) of generating y=1y = 1 givenx=xx =x

LNLL=logP(yx;θ)=1y=1logp1y=0log(1p)=ylogfθ(x)(1y)log(1fθ(x)) \begin{aligned} {L}_{NLL}=-logP(y|x;\theta)={-1}_{y=1}{logp-1}_{y=0}log(1-p)\\ =-ylog{f}_{\theta}(x)-(1-y)log(1-{f}_{\theta}(x)) \end{aligned}

3.1. 分类交叉熵损失(Categorical Cross-Entropy Loss)

分类交叉熵损失也被称为负对数似然(negative log likelihood)。这是一种用于解决分类问题的流行的损失函数,可用于测量两种概率分布(通常是真实标签和预测标签)之间的相似性。它可用 L=(ylog(yprediction))L = -\sum(y * \log(y_{prediction})) 表示,其中 y 是真实标签的概率分布(通常是一个one-hot vector),ypredictiony_{prediction} 是预测标签的概率分布,通常来自于一个 softmax。

3.2. Tukeys Loss

Robust Optimization for Deep Regression

TukeysBiweight

3.3. Dice Loss

常用于图像分割任务 Pytorch实现

4. Perceptual Loss

感知损失:可以将卷积神经网络提取出的feature,作为目标函数的一部分,通过比较待生成的图片经过CNN的feature值与目标图片经过CNN的feature值,使得待生成的图片与目标图片在语义上更加相似(相对于Pixel级别的损失函数)。

5. Focal loss

看ICCV那篇focal loss的论文《Focal Loss for Dense Object Detection》.

不过这个pytorch版detectron还没实现,官方Detectron是集成在Caffe2里。可参考Pytorch实现

Loss(x,class)=α(1softmax(x)[class])γlog(softmax(x)[class]) Loss(x, class) = - \alpha (1-softmax(x)_{[class]})^\gamma \log(softmax(x)_{[class]})

def focal_loss(inputs, targets):
    gamma = 2
    N = inputs.size(0)
    C = inputs.size(1)
    P = F.softmax(inputs) # softmax(x)

    class_mask = inputs.data.new(N, C).fill_(0)
    class_mask = Variable(class_mask)
    ids = targets.view(-1, 1)
    class_mask.scatter_(1, ids, 1.)
    # print(class_mask)

    probs = (P * class_mask).sum(1).view(-1, 1)# softmax(x)_class

    log_p = probs.log()
    # print('probs size= {}'.format(probs.size()))
    # print(probs)

    batch_loss = -(torch.pow((1 - probs), gamma)) * log_p
    # print('-----bacth_loss------')
    # print(batch_loss)

    loss = batch_loss.mean()

    return loss
  • α\alpha(1D Tensor, Variable) : the scalar factor for this criterion
  • γ\gamma(float, double) : γ>0\gamma > 0; reduces the relative loss for well-classified examples (p > .5), putting more focus on hard, misclassified examples
  • size_average(bool): By default, the losses are averaged over observations for each minibatch. However, if the field size_average is set to False, the losses are instead summed for each minibatch.

5.1. Huber Loss

loss(x,y)=1nizizi={0.5(xiyi)2,if xiyi<1xiyi0.5,otherwise  \begin{aligned} \text{loss}(x, y) = \frac{1}{n} \sum_{i} z_{i} \\ z_{i} = \begin{cases} 0.5 (x_i - y_i)^2, & \text{if } |x_i - y_i| < 1 \\ |x_i - y_i| - 0.5, & \text{otherwise } \end{cases} \end{aligned}

機器/深度學習: 損失函數(loss function)- Huber Loss和 Focal loss

results matching ""

    No results matching ""