<s>
Perceiver	B-General_Concept
is	O
a	O
transformer	B-Algorithm
adapted	O
to	O
be	O
able	O
to	O
process	O
non-textual	O
data	O
,	O
such	O
as	O
images	O
,	O
sounds	O
and	O
video	O
,	O
and	O
spatial	O
data	O
.	O
</s>
<s>
Transformers	B-Algorithm
underlie	O
other	O
notable	O
systems	O
such	O
as	O
BERT	B-General_Concept
and	O
GPT-3	B-General_Concept
,	O
which	O
preceded	O
Perceiver	B-General_Concept
.	O
</s>
<s>
It	O
adopts	O
an	O
asymmetric	O
attention	B-General_Concept
mechanism	I-General_Concept
to	O
distill	O
inputs	O
into	O
a	O
latent	O
bottleneck	O
,	O
allowing	O
it	O
to	O
learn	O
from	O
large	O
amounts	O
of	O
heterogeneous	O
data	O
.	O
</s>
<s>
Perceiver	B-General_Concept
matches	O
or	O
outperforms	O
specialized	O
models	O
on	O
classification	O
tasks	O
.	O
</s>
<s>
Perceiver	B-General_Concept
was	O
introduced	O
in	O
June	O
2021	O
by	O
DeepMind	B-Application
.	O
</s>
<s>
It	O
was	O
followed	O
by	O
Perceiver	B-General_Concept
IO	I-General_Concept
in	O
August	O
2021	O
.	O
</s>
<s>
Perceiver	B-General_Concept
is	O
designed	O
without	O
modality-specific	O
elements	O
.	O
</s>
<s>
It	O
uses	O
a	O
small	O
set	O
of	O
latent	O
units	O
that	O
forms	O
an	O
attention	B-General_Concept
bottleneck	O
through	O
which	O
the	O
inputs	O
must	O
pass	O
.	O
</s>
<s>
One	O
benefit	O
is	O
to	O
eliminate	O
the	O
quadratic	O
scaling	O
problem	O
found	O
in	O
early	O
transformers	B-Algorithm
.	O
</s>
<s>
Earlier	O
work	O
used	O
custom	O
feature	B-Algorithm
extractors	I-Algorithm
for	O
each	O
modality	B-General_Concept
.	O
</s>
<s>
It	O
associates	O
position	O
and	O
modality-specific	O
features	O
with	O
every	O
input	O
element	O
(	O
e.g.	O
</s>
<s>
every	O
pixel	B-Algorithm
,	O
or	O
audio	O
sample	O
)	O
.	O
</s>
<s>
These	O
features	O
can	O
be	O
learned	B-General_Concept
or	O
constructed	B-General_Concept
using	O
high-fidelity	O
Fourier	B-Algorithm
features	O
.	O
</s>
<s>
Perceiver	B-General_Concept
uses	O
cross-attention	O
to	O
produce	O
linear	O
complexity	O
layers	O
and	O
to	O
detach	O
network	O
depth	O
from	O
input	O
size	O
.	O
</s>
<s>
A	O
cross-attention	O
module	O
maps	O
a	O
(	O
larger	O
)	O
byte	O
array	O
(	O
e.g.	O
,	O
a	O
pixel	B-Algorithm
array	O
)	O
and	O
a	O
latent	O
array	O
(	O
smaller	O
)	O
to	O
another	O
latent	O
array	O
,	O
reducing	B-Algorithm
dimensionality	I-Algorithm
.	O
</s>
<s>
A	O
transformer	B-Algorithm
tower	O
maps	O
one	O
latent	O
array	O
to	O
another	O
latent	O
array	O
,	O
which	O
is	O
used	O
to	O
query	O
the	O
input	O
again	O
.	O
</s>
<s>
Both	O
components	O
user	O
query-key-value	O
(	O
QKV	O
)	O
attention	B-General_Concept
.	O
</s>
<s>
QKV	O
attention	B-General_Concept
applies	O
query	O
,	O
key	O
,	O
and	O
value	O
networks	O
,	O
which	O
are	O
typically	O
multilayer	B-Algorithm
perceptrons	I-Algorithm
–	O
to	O
each	O
element	O
of	O
an	O
input	O
array	O
,	O
producing	O
three	O
arrays	O
that	O
preserve	O
the	O
index	O
dimensionality	O
(	O
or	O
sequence	O
length	O
)	O
of	O
their	O
inputs	O
.	O
</s>
<s>
Perceiver	B-General_Concept
IO	I-General_Concept
can	O
flexibly	O
query	O
the	O
model	O
's	O
latent	O
space	O
to	O
produce	O
outputs	O
of	O
arbitrary	O
size	O
and	O
semantics	O
.	O
</s>
<s>
It	O
achieves	O
results	O
on	O
tasks	O
with	O
structured	O
output	O
spaces	O
,	O
such	O
as	O
natural	B-General_Concept
language	I-General_Concept
and	O
visual	B-Application
understanding	O
,	O
StarCraft	B-Application
II	I-Application
,	O
and	O
multi-tasking	O
.	O
</s>
<s>
Perceiver	B-General_Concept
IO	I-General_Concept
matches	O
a	O
Transformer-based	O
BERT	B-General_Concept
baseline	O
on	O
the	O
GLUE	O
language	O
benchmark	O
without	O
the	O
need	O
for	O
input	O
tokenization	O
and	O
achieves	O
state-of-the-art	O
performance	O
on	O
Sintel	O
optical	O
flow	O
estimation	O
.	O
</s>
<s>
For	O
example	O
to	O
predict	O
optical	O
flow	O
on	O
one	O
pixel	B-Algorithm
a	O
query	O
would	O
attend	O
using	O
the	O
pixel	B-Algorithm
’s	O
xy	O
coordinates	O
plus	O
an	O
optical	O
flow	O
task	O
embedding	O
to	O
produce	O
a	O
single	O
flow	O
vector	O
.	O
</s>
<s>
Perceiver	B-General_Concept
's	O
performance	O
is	O
comparable	O
to	O
ResNet-50	O
and	O
ViT	B-Algorithm
on	O
ImageNet	B-General_Concept
without	O
2D	O
convolutions	B-Architecture
.	O
</s>
<s>
It	O
attends	O
to	O
50,000	O
pixels	B-Algorithm
.	O
</s>
<s>
It	O
is	O
competitive	O
in	O
all	O
modalities	B-General_Concept
in	O
AudioSet	O
.	O
</s>
