[Paper Review] 19. GAN Dissection : Visualizing and Understanding GANs
Contents
- Abstract
- Introduction
- Method
- Characterizing units by Dissection
- Measuring Causal Relationships using Intervention
0. Abstract
present an analytic framework to visualize & understand GANs at
- unit / object / scene level
Step 1 ) identify a group of interpretable units
- that are closely related to object concepts,
- using a segmentation-based network dissection method
Step 2 ) quantify the causal effect of interpretable units
- by measuring the ability of interventions to control objects in the output
1. Introduction
General method for visualizing & understanding GANs
- at different level of abstractions
- from each neuron / object / contextual relationship
2. Method
-
analyze how objects such as trees are encoded by internal representations of GAN generator
-
notation
- GAN : \(G : \mathbf{z} \rightarrow \mathbf{x}\) …. where \(\mathrm{x} \in \mathbb{R}^{H \times W \times 3}\)
- \(\mathrm{x}=f(\mathbf{r})=f(h(\mathbf{z}))=G(\mathbf{z})\).
-
\(\mathbf{r}\) has necessary data to produce images!
Question
- Whether information about concept \(c\) is in \(\mathbf{r}\)? (X)
- How such information is encoded in \(\mathbf{r}\) ! (O)
-
\(\mathbf{r}_{\mathbb{U}, \mathrm{P}}=\left(\mathbf{r}_{\mathrm{U}, \mathrm{P}}, \mathbf{r}_{\overline{\mathrm{U}}, \mathrm{P}}\right)\).
- generation of object \(c\) at location \(P\) depends mainly on \(\mathbf{r}_{\mathrm{U}, \mathrm{P}}\)
- insensitive to other units \(\mathbf{r}_{\bar{U}, \mathrm{P}}\)
Structure of \(\mathbf{r}\) in 2 phases
-
Dissection
measuring agreement between individual units of \(\mathbf{r}\) & every class \(c\)
-
Intervention
for the represented classes (identified through 1),
identify causal sets of units & measure causal effects
( by forcing sets of units on/off )
(1) Characterizing units by Dissection
\(\mathbf{r}_{u, \mathbb{P}}\) : one-channel \(h \times w\) feature map of unit \(u\)
\(\rightarrow\) Q) does \(\mathbf{r}_{u, \mathbb{P}}\) encodes semantic class (ex.tree)?
(1) Select a universe of concepts \(c \in \mathbb{C}\), for which we have semantic segmentation \(s_c(x)\) for each class
(2) Then, quantify spatial agreement between
- 1) unit \(u\)’s thresholded feature map
- 2) concept \(c\)’s segmentation
with IOU measure
(2) Measuring Causal Relationships using Intervention
test whether a set of units \(U\) in \(\mathbf{r}\) cause the generation of \(c\) !
- via turning on/off the units of \(U\)
Decompose feature map \(\mathbf{r}\) into 2 parts : \(\left(\mathbf{r}_{\mathrm{U}, \mathrm{P}}, \mathrm{r}_{\overline{U, \mathrm{P}}}\right)\)
( where \(\mathrm{r}_{\overline{U, P}}\) : unforced components of \(\mathbf{r}\) )
Original Image :
- \(\mathbf{x}=G(\mathbf{z}) \equiv f(\mathbf{r}) \equiv f\left(\mathbf{r}_{\mathrm{U}, \mathrm{P}}, \mathbf{r}_{\overline{\mathrm{U}, \mathrm{P}}}\right)\).
Image with \(U\) ablated at pixels \(P\) :
- \[\mathbf{x}_{a}=f\left(\mathbf{0}, \mathbf{r}_{\overline{\mathrm{U}, \mathrm{P}}}\right)\]
Image with \(U\) inserted at pixels \(P\) :
- \(\mathbf{x}_{i}=f\left(\mathbf{k}, \mathbf{r}_{\overline{U, \mathrm{P}}}\right)\).
Object is caused by \(U\) if …
- the object appears in \(x_i\)
- the object disappears from \(x_a\)
This causality can be quantified by..
- comparing presence of trees in \(x_i\) & \(x_a\)
- average effects over all locations & images
ACE
- average causal effect (ACE) of units \(\mathrm{U}\) on the generation of on class \(c\)
- \(\delta_{\mathrm{U} \rightarrow c} \equiv \mathbb{E}_{\mathbf{z}, \mathrm{P}}\left[\mathbf{s}_{c}\left(\mathbf{x}_{i}\right)\right]-\mathbb{E}_{\mathbf{z}, \mathrm{P}}\left[\mathbf{s}_{c}\left(\mathbf{x}_{a}\right)\right]\).
