This page shows visualizations of some width-1 1-d convolutional filters from Google's lm_1b language model. Each column corresponds to one position in the filter, and shows the characters with the most positive weights. Use the checkbox in the bottom-right to also see the most negative weights (may be slow).

Below that are examples of words for which the filter emits the highest values. A filter's response is its maximum value over all substrings it sees in the word. So if a filter has high weights on 'c' in the first position, then 'a', then 't', it will assign equally high scores to 'cat', 'fatcat', 'concatenate', etc. The portion of the string in blue is the substring the filter is responding to.

'^' and '$' represent beginning and end of word markers, respectively. '_' is a padding character. Literal versions of those characters are escaped with a backslash.

Use the links at the top to see filters of other widths.

Check out my blog post here for a bit more context.

Filter 0 (bias = -0.19) #

x U
<BOS> y
S m
j T
5 D
7 u
9 P
\194 L
V z
4 H
6 d
Q K
8 c
F M
v A
' \195
X r
N O
0 t
b G

Non-zero for 7.9% of words.

Filter 1 (bias = -0.25) #

N i
. p
<BOS> g
Q m
F k
/ P
K Y
B V
r G
9 T
8 d
M l
L -
c h
Z 1
o H
f v
A b
R z
E J

Non-zero for 3.6% of words.

Filter 2 (bias = -0.46) #

z h
<BOS> H
Q y
v F
U M
I j
\194 4
G e
K u
s f
l i
a 3
w 1
<EOS> n
O g
' r
c N
. q
R x
/ 8

Non-zero for 9.8% of words.

Filter 3 (bias = -0.25) #

W j
w D
q P
" l
i R
a d
y r
K F
X J
2 L
B u
4 C
O \195
^ -
, 0
G
s
g
z
U

Non-zero for 6.0% of words.

Filter 4 (bias = -0.40) #

. i
L P
A p
Q k
<BOS> J
/ f
g 3
a K
\194 B
l 1
q ,
m 9
N R
Z T
u 0
c 2
d j
- I
b O
h 4

Non-zero for 4.3% of words.

Filter 5 (bias = -0.08) #

1 m
W f
q L
I F
g M
R l
k K
v .
Y /
i y
2 D
9 e
a j
p U
0 Z
7 t
" s
h P
3 <BOS>
- S

Non-zero for 0.0% of words.

Filter 6 (bias = -0.25) #

C v
G q
P e
s b
1 E
D x
z f
R w
n W
/ B
c -
A .
5 N
7 u
, h
g <BOS>
I a
V M
Z T
U \163

Non-zero for 3.8% of words.

Filter 7 (bias = -0.32) #

. P
- F
' J
<BOS> 2
Y B
x i
m 3
Q j
g 5
o E
" 1
b K
\194 4
h 0
q 6
r I
a D
/ e
l 9
u T

Non-zero for 2.8% of words.

Filter 8 (bias = -0.34) #

. i
F p
Q P
<BOS> J
L O
N 1
A k
/ G
Z o
\194 T
q 3
b Y
c 0
' K
n 2
X R
x I
D z
" -
a \195

Non-zero for 6.5% of words.

Filter 9 (bias = -0.12) #

C -
7 e
A w
V v
Q E
1 o
, J
H f
s b
4 <BOS>
R m
U q
S .
F T
n M
8 N
k \195
Y K
Z l
/ B

Non-zero for 2.6% of words.

Filter 10 (bias = -0.48) #

4 m
7 u
5 .
3 -
6 U
2 z
8 v
X t
9 c
1 o
V T
S s
F '
W r
Q d
0 w
H b
, l
I y
i k

Non-zero for 10.2% of words.

Filter 11 (bias = -0.36) #

. p
N i
L P
/ k
Q T
\194 J
" I
s V
<BOS> 1
u j
' 0
y G
x B
c g
Z \195
A 2
o z
<EOS> v
F 3
8 E

Non-zero for 3.1% of words.

Filter 12 (bias = -0.11) #

Y F
X f
G e
W t
l s
V d
\194 c
H j
o E
<EOS> n
J p
^ P
L I
T C
\195 \163
b x
m S
z ,
/ k
" r

Non-zero for 0.2% of words.

Filter 13 (bias = -0.13) #

<BOS> o
j y
F c
Q m
I K
V U
7 w
X O
4 u
6 z
\194 a
P T
C G
H v
^ x
5 W
9 h
R p
l f
J i

Non-zero for 4.0% of words.

Filter 14 (bias = -0.16) #

. P
<BOS> R
\194 p
L i
q k
l U
A O
X r
t G
v 3
e J
m K
/ s
h \195
F o
D 1
Q 9
j f
M y
N C

Non-zero for 0.6% of words.

Filter 15 (bias = -0.34) #

H x
I v
A z
1 c
N p
Q o
4 G
2 f
7 m
X b
3 s
F '
q K
Z -
/ g
r 0
u S
t <EOS>
^ k
j .

Non-zero for 8.3% of words.

Filter 16 (bias = -0.48) #

j y
<BOS> U
7 m
J a
0 K
R w
\194 f
V z
9 s
4 .
Y p
I L
6 A
X c
5 x
D W
g o
S ,
H O
- B

Non-zero for 7.3% of words.

Filter 17 (bias = -0.08) #

P x
M .
D v
O a
j b
I u
X h
C s
Q -
T z
F w
H m
3 q
1 l
r o
K B
2 <EOS>
7 A
Z L
5 '

Non-zero for 0.0% of words.

Filter 18 (bias = -0.21) #

L k
. p
Z t
/ i
8 T
D I
5 v
6 r
A R
l P
F '
X C
G q
N -
K O
m W
<EOS> \163
M w
U ,
7 f

Non-zero for 2.0% of words.

Filter 19 (bias = -0.31) #

Z t
X u
L k
K -
6 R
8 r
5 h
Q s
2 v
G T
/ j
V o
W i
7 '
3 f
4 S
^ d
9 g
m I
<EOS> C

Non-zero for 7.5% of words.

Filter 20 (bias = -0.33) #

N k
E v
O p
L m
K g
Q '
U V
2 C
3 i
8 n
/ x
" c
9 j
A b
e -
I h
a t
W l
r T
o \194

Non-zero for 7.1% of words.

Filter 21 (bias = -0.45) #

. i
A J
Q K
c P
g 3
' p
h O
b o
C f
\194 2
l 1
t E
L 9
F W
a M
q 6
<BOS> 4
/ 5
d w
r ,

Non-zero for 2.9% of words.

Filter 22 (bias = -0.41) #

. s
q U
Q G
<BOS> J
\194 i
X K
v P
r S
a E
b 3
N z
' O
t 5
n L
" o
h f
W R
H 0
\163 M
- 4

Non-zero for 1.6% of words.

Filter 23 (bias = -0.09) #

<BOS> u
x U
. T
X k
Q H
5 i
7 h
\194 R
6 y
8 t
G B
l r
^ m
2 M
p Y
0 P
/ A
9 \195
V 1
Z D

Non-zero for 3.7% of words.

Filter 24 (bias = -0.24) #

F -
S o
Q u
5 \195
C J
8 w
4 v
7 T
6 r
, Y
X b
/ R
L z
\194 i
f k
D m
s g
e q
j p
2 a

Non-zero for 8.5% of words.

Filter 25 (bias = -0.21) #

W u
x -
X D
B d
V r
S R
K j
<BOS> U
\194 J
Q P
5 \195
" g
6 s
, I
b E
w T
p L
Y H
4 1
q o

Non-zero for 9.0% of words.

Filter 26 (bias = -0.36) #

B m
N d
E c
<BOS> p
9 y
J g
R .
S -
3 D
4 '
I G
F C
2 z
5 v
K l
W n
, \163
7 L
8 x
6 T

Non-zero for 14.9% of words.

Filter 27 (bias = -0.17) #

I x
<BOS> h
Q v
/ b
, m
2 c
P g
N k
R y
O f
n B
D G
H Y
t V
A p
^ M
C T
1 o
7 S
X 0

Non-zero for 1.1% of words.

Filter 28 (bias = -0.38) #

A -
S v
Q u
<BOS> d
N m
E c
O D
W P
I n
X '
F U
K T
, J
5 C
B y
r p
e z
t M
4 i
a 0

Non-zero for 10.6% of words.

Filter 29 (bias = -0.33) #

k L
T l
' D
U o
r .
W 5
B J
" d
i /
V 6
R -
t j
C 0
b G
p 7
v N
I K
Q X
\163 <EOS>
H n

Non-zero for 4.3% of words.

Filter 30 (bias = -0.51) #

. P
- J
' i
\194 K
<BOS> k
n B
m 3
l p
d U
t E
q R
g 2
/ 9
h 1
x G
c T
A 0
Q O
L \195
v 4

Non-zero for 2.9% of words.

Filter 31 (bias = -0.33) #

O n
W u
S v
E F
G k
Y C
" j
Q f
o R
i J
w Z
p -
y b
A M
2 P
K B
a D
, m
^ '
t 9

Non-zero for 6.9% of words.