1 of 11

Understanding neural networks from their weights

Thomas Dooms

2 of 11

Baseline

Your brilliant idea

time

metric

3 of 11

Understanding neural networks from their weights

Thomas Dooms

4 of 11

The weights in ordinary neural networks

are not meaningful

-1.0

1.0

-1.0

1.0

Input 1

Input 2

0.0

1.0

Input 3

-1.0

0.0

0.0

ReLU

Output 3

Output 1

Output 2

5 of 11

Most activation functions can represent

feature interactions of any degree

Input 1

Input 2

Input 3

Input 4

6 of 11

Activation functions can be replaced with

a linear gate without a drop in accuracy

Output

ReLU

Linear

Input

Output

Linear

Input

Linear

x

7 of 11

Each output is described by

a sum of pairwise feature interactions

-1.0

1.0

-1.0

1.0

Input 1

Input 2

Input 1

Input 2

8 of 11

We can use PCA to reveal

the important components

9 of 11

These components can be used

to fully decompose the model

Eigenvalues

Eigenvectors

10 of 11

Analyzing model weights can reveal issues

including learning unexpected patterns

Test accuracy: 98.1%

Test accuracy: 98.2%

11 of 11

Understanding neural networks from their weights

Thomas Dooms