论文阅读|INTRA-CLASS UNCERTAINTY LOSS FUNCTION FOR CLASSIFICATION

主要观点

• 提出基于高斯分布的损失函数，以处理类内不确定性问题
• 高斯均值相当于类中心，不确定性用方差描述
• 类似于类间margin，提出类内margin

原理阐述

类样本不平衡会导致极坐标下的类分类变得困难

L-GM Loss:
$$\begin{array}{rl}{\mathcal{L}}_{L-GM}=& -\frac{1}{N}\sum _{i=1}^N\log \frac{e^{-d_{z_i}(1+\alpha )}}{{\sum }_m^M{e}^{-d_m\left(1+R\left(m=z_i\right)\alpha \right)}}\\ & +\lambda \left(d_{z_i}+\frac{1}{2}\log \left|{\Lambda }_{z_i}\right|\right)\end{array}$$
其中，$d_m=\frac{1}{2}{\left(x_i-{\mu }_m\right)}^T{\Lambda }_m^{-1}\left(x_i-{\mu }_m\right),m\in [1,M]$

Center Loss:
$${\mathcal{L}}_{\text{Center }}=-\frac{1}{N}\sum _i^N\log \frac{e^{w_{y_i}^T{x}_i}}{{\sum }_{k=0}^K{e}^{w_k^T{x}_i}}+\frac{\lambda }{2}\sum _i^N{\Vert x_i-c_{y_i}\Vert }_2^2$$

ICU Loss:
$$\begin{array}{rl}{\mathcal{L}}_{cls}& =-\frac{1}{N}\sum _{i=1}^N\log p\left(z_i\mid x_i\right)\\ & =-\frac{1}{N}\sum _{i=1}^N\log \frac{e^{-d_{z_i}}}{{\Sigma }_k^K{e}^{-d_k}}+\lambda {\mathcal{L}}_{reg}\end{array}$$
其中，$d_k=\frac{1}{2}[{\left(x_i-{\mu }_k\right)}^T{\Sigma }_k^{-1}\left(x_i-{\mu }_k\right)+\ln \mid {\Sigma }_k\Vert ,k\in [1,K]$
$$\lambda {\mathcal{L}}_{\text{reg }}=\sum _{k=1}^K{\lambda }_1{\left|{\mu }_k-{\bar{\mu }}_{N_k}\right|}^2+{\lambda }_2{\left|{\sigma }_k^2-{\bar{\sigma }}_{N_k}^2\right|}^2$$
ICU Loss相当于Center Loss + L-GM Loss + 方差规整

现在再加上类内外的margin项，得到：
$${\mathcal{L}}_{ICU}=-\frac{1}{N}\sum _{i=1}^N\log \frac{e^{-d_{z_i}(1+\alpha )}}{{\Sigma }_{k,k\ne z_i}^K{e}^{-d_k}+e^{-d_{z_i}(1+\alpha )}}$$
$$d_{z_i}=\frac{1}{2}\left[{\left(x_i-{\mu }_{z_i}\right)}^T{\Sigma }_{z_i}^{-1}\left(x_i-{\mu }_{z_i}\right)+\ln (1+\gamma )\left|{\Sigma }_{z_i}\right|\right]$$
$$d_k=\frac{1}{2}\left[{\left(x_i-{\mu }_k\right)}^T{\Sigma }_k^{-1}\left(x_i-{\mu }_k\right)+\ln \left|{\Sigma }_k\right|\right],k\in [1,K]$$
其中，$\alpha$控制类外margin，$\gamma$控制类内margin