+ All Categories
Home > Documents > What Do Gates Learn: Visualization & Analysis

What Do Gates Learn: Visualization & Analysis

Date post: 18-Feb-2016
Category:
Upload: zhengbo-lee
View: 8 times
Download: 1 times
Share this document with a friend
Description:
This presentation slides includes motivation and methods in visualizing neural networks, combined with analysis and plan for future work.
Popular Tags:
18
1 What do gates learn? Visualization & Analysis Zhengbo Li [email protected] Shanghai Jiao Tong University
Transcript
Page 1: What Do Gates Learn: Visualization & Analysis

1

What do gates learn?Visualization & Analysis

Zhengbo [email protected]

Shanghai Jiao Tong University

Page 2: What Do Gates Learn: Visualization & Analysis

2

Outline

1. Why we want to visualize?

2. How can we visualize?

3. Experiment results

4. Future work

Page 3: What Do Gates Learn: Visualization & Analysis

3

Why visualize?

What do gates learn? Weights! Correct but meaningless for human beings.Weight matrix between two adjacent layers:

Page 4: What Do Gates Learn: Visualization & Analysis

4

Why visualize?

We have much more ability in dealing with pictures than numbers

We have powerful eyes: over 500 million pixels

Page 5: What Do Gates Learn: Visualization & Analysis

5

About data & network

Data: 1000 random rectangles on 10 by 10 gridExample:

Structure [100, 10, 100], fully connected

Page 6: What Do Gates Learn: Visualization & Analysis

6

How to visualize

If Gate learns some property, we may expect relationship between property and activationfor each gate:

Page 7: What Do Gates Learn: Visualization & Analysis

7

How to visualize

If Gate learns some property, we may expect relationship between property and activationfor each gate:

Page 8: What Do Gates Learn: Visualization & Analysis

8

Experiment Results

For 10 hidden gates:

Page 9: What Do Gates Learn: Visualization & Analysis

9

What can we find?

For one of the gates, it seems that there is a linear relationship between the activation and area, we can call it linear gate.

For many independent experiments, we can always find exactly one linear gate.

It may help if we visualize while training process

Page 10: What Do Gates Learn: Visualization & Analysis

10

Start of training

Page 11: What Do Gates Learn: Visualization & Analysis

11

Middle of training

Page 12: What Do Gates Learn: Visualization & Analysis

12

End of training

Page 13: What Do Gates Learn: Visualization & Analysis

13

Some explanation

Auto encoder → dimension reduction

Less redundancy between reduced dimensions

Linear gate + linear gate = linear gate, thus it is better to have only one linear gate!

Page 14: What Do Gates Learn: Visualization & Analysis

14

Visualization method 2

If a gate learns some property, the gate is activated when a input rectangle has this property.

For each gate, add the rectangles that can cause high activation for this gate.

Page 15: What Do Gates Learn: Visualization & Analysis

15

Experiment Result 2

Page 16: What Do Gates Learn: Visualization & Analysis

16

Explanation 2

Gates learn about rectangles by being sensitive to different areas of the gird.

Page 17: What Do Gates Learn: Visualization & Analysis

17

Future work

1. Find more visualization methods.

2. Try different network structures

3. Try different set of input (circles, diamonds)

4. Formalize the problem and try to prove something.

Page 18: What Do Gates Learn: Visualization & Analysis

18

Thank you for listening!

Comments & Suggestions?


Recommended