让机器用人的方式识别图像[codes]

论文地址

This looks like that: deep learning for interpretable image recognition

prototype layer

prototype layer对应的公式如下:
$g_{p_j}=\max_{ \vec{z} \in patch(z)} -log( || \vec{z} - p_j ||^2_2 + \epsilon )$
其中max层可以用现成的maxpool层来实现,下面重点讨论新的层实现
$f_{z}= -log( || \vec{z} - p_j ||^2_2 + \epsilon )$
这个层的输入输出的通道数可以不一致,但是每个通道内的尺寸需要一致. 为了后续讨论方便,约定以下符号
* in_data: (batchSize, inChNum, height, width), 输入数据,即公式中的$z$
* weight: (outChNum, inChNum, 3, 3), filter权重, 此处假设采用3x3的filter
* out_data: (batchSize, outChNum, height, width), 输出数据,迹公式中的$f(z)$

in_data (one sample)

$z_{00}$	$z_{01}$	$z_{02}$
$z_{10}$	$z_{11}$	$z_{12}$
$z_{20}$	$z_{21}$	$z_{22}$
$z_{30}$	$z_{31}$	$z_{32}$

weight (one sample and one output channel)

$w_{00}$	$w_{01}$