1. 1. Contribution
2. 2. Network
   1. 2.1 ITSA Loss

1. Contribution

又是一点凑了三点。

1. learning feature representations that are less sensitive to input variations
2. novel loss function that enables us to minimize the Fisher information, without computing the second-order derivatives.
3. can be used in training models for non-geometry based vision problems such as semantic segmentation

2. Network

2.1 ITSA Loss

看图就够了，不知道怎么凑出来最后的$L_{FI}$函数的😀。



$f_\theta$是特征提取网络