U.S. patent application number 16/751167 was filed with the patent office on 2020-07-23 for generating neural network outputs using insertion operations.
The applicant listed for this patent is Google LLC. Invention is credited to William Chan, Jamie Ryan Kiros, Mitchell Thomas Stern, Jakob D. Uszkoreit.
Application Number | 20200234011 16/751167 |
Document ID | / |
Family ID | 69743906 |
Filed Date | 2020-07-23 |
![](/patent/app/20200234011/US20200234011A1-20200723-D00000.png)
![](/patent/app/20200234011/US20200234011A1-20200723-D00001.png)
![](/patent/app/20200234011/US20200234011A1-20200723-D00002.png)
![](/patent/app/20200234011/US20200234011A1-20200723-D00003.png)
![](/patent/app/20200234011/US20200234011A1-20200723-D00004.png)
![](/patent/app/20200234011/US20200234011A1-20200723-M00001.png)
![](/patent/app/20200234011/US20200234011A1-20200723-M00002.png)
![](/patent/app/20200234011/US20200234011A1-20200723-M00003.png)
United States Patent
Application |
20200234011 |
Kind Code |
A1 |
Uszkoreit; Jakob D. ; et
al. |
July 23, 2020 |
GENERATING NEURAL NETWORK OUTPUTS USING INSERTION OPERATIONS
Abstract
Methods, systems, and apparatus, including computer programs
encoded on computer storage media, for generating network outputs
using insertion operations.
Inventors: |
Uszkoreit; Jakob D.;
(Portola Valley, CA) ; Stern; Mitchell Thomas;
(Berkeley, CA) ; Kiros; Jamie Ryan; (Toronto,
CA) ; Chan; William; (Markham, CA) |
|
Applicant: |
Name |
City |
State |
Country |
Type |
Google LLC |
Mountain View |
CA |
US |
|
|
Family ID: |
69743906 |
Appl. No.: |
16/751167 |
Filed: |
January 23, 2020 |
Related U.S. Patent Documents
|
|
|
|
|
|
Application
Number |
Filing Date |
Patent Number |
|
|
62796038 |
Jan 23, 2019 |
|
|
|
62815908 |
Mar 8, 2019 |
|
|
|
Current U.S.
Class: |
1/1 |
Current CPC
Class: |
G06F 40/44 20200101;
G06N 3/082 20130101; G06N 5/04 20130101; G06N 3/0445 20130101; G06N
3/08 20130101; G06N 3/0454 20130101 |
International
Class: |
G06F 40/44 20060101
G06F040/44; G06N 5/04 20060101 G06N005/04; G06N 3/04 20060101
G06N003/04; G06N 3/08 20060101 G06N003/08 |
Claims
1. A method performed by one or more computers, the method
comprising: receiving a network input; and generating a network
output from the network input, wherein the network output comprises
a plurality of outputs from a vocabulary of outputs arranged
according to an output order, the generating comprising, at each of
a plurality of generation time steps: identifying a current partial
network output that has already been generated as of the generation
time step, the current partial network output comprising zero or
more outputs from the vocabulary of outputs arranged according to a
partial output order; generating, using a decoder neural network
conditioned on (i) at least a portion of the network input and (ii)
any outputs in the current partial network output, a decoder output
that defines, for each of a plurality of insertion locations, a
respective score distribution over the vocabulary of outputs,
wherein each insertion location is a different new location in the
partial output order at which there is no output in the current
partial network output; selecting, using the decoder output, one or
more of the insertion locations and, for each selected insertion
location, an inserted output from the vocabulary; and generating a
new partial network output that comprises (i) the zero or more
outputs in the current partial network output and (ii) for each
selected insertion location, the inserted output from the
vocabulary inserted at the corresponding new location in the
partial output order.
2. The method of claim 1, wherein the decoder neural network is an
attention-based neural network that is configured to generate the
decoder output by applying an attention mechanism over an encoded
representation of the network input and a self-attention mechanism
over the outputs in the current partial network output.
3. The method of claim 2, wherein generating the decoder output
using the decoder neural network comprises: generating a decoder
input that includes the encoded representation of the network input
and the outputs in the current partial network output arranged
according to the partial output order.
4. The method of claim 3, wherein generating the decoder input
further comprises adding two marker outputs to the current partial
network output, wherein the decoder neural network is configured to
generate a respective representation vector for each location in
the partial output order after the two marker outputs have been
added, and wherein generating the decoder output comprises:
generating a respective slot representation for each insertion
location by concatenating the representation vectors for each
adjacent pair of locations in the partial output order; and
generating a score distribution for each insertion location from at
least the slot representation for the insertion location.
5. The method of claim 1, wherein the vocabulary includes an
end-of-sequence token.
6. The method of claim 5, wherein selecting, using the decoder
output, one or more of the insertion locations and, for each
selected insertion location, an inserted output from the vocabulary
comprises: determining that an insertion location--output
combination with a highest score across all insertion
location--output combinations does not include the end-of-sequence
token; and in response, selecting only the insertion
location--output combination with a highest score across all
insertion location--output combinations.
7. The method of claim 5, wherein selecting, using the decoder
output, one or more of the insertion locations and, for each
selected insertion location, an inserted output from the vocabulary
comprises: determining that there is at least one insertion
location for which the output with the highest score is not the
end-of-sequence token; and in response, selecting only an insertion
location--output combination with a highest score across all
insertion location--output combinations that include an insertion
location for which the output with the highest score is not the
end-of-sequence token.
8. The method of claim 5, wherein selecting, using the decoder
output, one or more of the insertion locations and, for each
selected insertion location, an inserted output from the vocabulary
comprises: identifying, from the decoder output and for each
insertion location, an output that has a highest score for the
insertion location; determining that there is at least one
insertion location for which the output with the highest score is
not the end-of-sequence token; and in response, selecting each
insertion location for which the output with the highest score is
not the end-of-sequence token and the corresponding output that has
the highest score for the insertion location.
9. The method of claim 1, wherein the decoder neural network is
configured to generate a respective slot representation for each
insertion location.
10. The method of claim 9, wherein generating the decoder output
comprises: projecting a decoder hidden state matrix generated from
the slot representations using a projection matrix to generate a
content-location logit matrix; flattening the content-location
logit matrix into a content-location logit vector; and applying a
softmax over the content-location logit vector to generate a
probability distribution over all insertion location--output
combinations.
11. The method of claim 9, wherein generating the decoder output
comprises: generating a respective probability for each location by
applying a softmax to a product of a decoder hidden state matrix
generated from the slot representations and a learned query vector;
for each location: projecting the slot representation for the
location into a score vector that includes a respective score for
each output in the vocabulary using a projection matrix; applying a
softmax over the score vector to generate an initial probability
for each output in the vocabulary; and multiplying each initial
probability by the probability for the location to generate a final
probability for each output in the vocabulary.
12. The method of claim 9, wherein generating the decoder output
comprises: generating a context vector by applying max pooling over
the slot representations; generating a bias vector from the context
vector that includes a respective bias value for each output in the
vocabulary; and generating the decoder output from the bias vector
and the slot representations.
13. One or more non-transitory computer-readable storage media
storing instructions that when executed by one or more computers
cause the one or more computers to perform operations comprising:
receiving a network input; and generating a network output from the
network input, wherein the network output comprises a plurality of
outputs from a vocabulary of outputs arranged according to an
output order, the generating comprising, at each of a plurality of
generation time steps: identifying a current partial network output
that has already been generated as of the generation time step, the
current partial network output comprising zero or more outputs from
the vocabulary of outputs arranged according to a partial output
order; generating, using a decoder neural network conditioned on
(i) at least a portion of the network input and (ii) any outputs in
the current partial network output, a decoder output that defines,
for each of a plurality of insertion locations, a respective score
distribution over the vocabulary of outputs, wherein each insertion
location is a different new location in the partial output order at
which there is no output in the current partial network output;
selecting, using the decoder output, one or more of the insertion
locations and, for each selected insertion location, an inserted
output from the vocabulary; and generating a new partial network
output that comprises (i) the zero or more outputs in the current
partial network output and (ii) for each selected insertion
location, the inserted output from the vocabulary inserted at the
corresponding new location in the partial output order.
14. A system comprising one or more computers and one or more
storage devices storing instructions that when executed by one or
more computers cause the one or more computers to perform
operations comprising: receiving a network input; and generating a
network output from the network input, wherein the network output
comprises a plurality of outputs from a vocabulary of outputs
arranged according to an output order, the generating comprising,
at each of a plurality of generation time steps: identifying a
current partial network output that has already been generated as
of the generation time step, the current partial network output
comprising zero or more outputs from the vocabulary of outputs
arranged according to a partial output order; generating, using a
decoder neural network conditioned on (i) at least a portion of the
network input and (ii) any outputs in the current partial network
output, a decoder output that defines, for each of a plurality of
insertion locations, a respective score distribution over the
vocabulary of outputs, wherein each insertion location is a
different new location in the partial output order at which there
is no output in the current partial network output; selecting,
using the decoder output, one or more of the insertion locations
and, for each selected insertion location, an inserted output from
the vocabulary; and generating a new partial network output that
comprises (i) the zero or more outputs in the current partial
network output and (ii) for each selected insertion location, the
inserted output from the vocabulary inserted at the corresponding
new location in the partial output order.
15. The system of claim 14, wherein the decoder neural network is
an attention-based neural network that is configured to generate
the decoder output by applying an attention mechanism over an
encoded representation of the network input and a self-attention
mechanism over the outputs in the current partial network
output.
16. The system of claim 15, wherein generating the decoder output
using the decoder neural network comprises: generating a decoder
input that includes the encoded representation of the network input
and the outputs in the current partial network output arranged
according to the partial output order.
17. The system of claim 16, wherein generating the decoder input
further comprises adding two marker outputs to the current partial
network output, wherein the decoder neural network is configured to
generate a respective representation vector for each location in
the partial output order after the two marker outputs have been
added, and wherein generating the decoder output comprises:
generating a respective slot representation for each insertion
location by concatenating the representation vectors for each
adjacent pair of locations in the partial output order; and
generating a score distribution for each insertion location from at
least the slot representation for the insertion location.
18. The system of claim 14, wherein the vocabulary includes an
end-of-sequence token.
19. The system of claim 18, wherein selecting, using the decoder
output, one or more of the insertion locations and, for each
selected insertion location, an inserted output from the vocabulary
comprises: determining that an insertion location--output
combination with a highest score across all insertion
location--output combinations does not include the end-of-sequence
token; and in response, selecting only the insertion
location--output combination with a highest score across all
insertion location--output combinations.
20. The system of claim 18, wherein selecting, using the decoder
output, one or more of the insertion locations and, for each
selected insertion location, an inserted output from the vocabulary
comprises: determining that there is at least one insertion
location for which the output with the highest score is not the
end-of-sequence token; and in response, selecting only an insertion
location--output combination with a highest score across all
insertion location--output combinations that include an insertion
location for which the output with the highest score is not the
end-of-sequence token.
Description
CROSS-REFERENCE TO RELATED APPLICATIONS
[0001] This application claims priority to U.S. Provisional
Application No. 62/796,038, filed on Jan. 23, 2019, and U.S.
Provisional Application No. 62/815,908, filed on Mar. 8, 2019. The
disclosures of the prior applications are considered part of and
are incorporated by reference in the disclosure of this
application.
BACKGROUND
[0002] This specification relates to generating outputs using
neural networks.
[0003] Neural networks are machine learning models that employ one
or more layers of nonlinear units to predict an output for a
received input. Some neural networks include one or more hidden
layers in addition to an output layer. The output of each hidden
layer is used as input to the next layer in the network, i.e., the
next hidden layer or the output layer. Each layer of the network
generates an output from a received input in accordance with
current values of a respective set of parameters.
SUMMARY
[0004] This specification describes a system implemented as
computer programs on one or more computers in one or more locations
that generates a network output that includes a respective output
from a vocabulary of outputs at each of multiple locations in an
output order. In some cases, the output is a one-dimensional
sequence, e.g., a sequence of text, while in other cases the output
is a higher-dimensional array, e.g., an image.
[0005] Particular embodiments of the subject matter described in
this specification can be implemented so as to realize one or more
of the following advantages.
[0006] Auto-regressive models have been shown to achieve
high-quality of performance on a variety of output generation
tasks, e.g., speech recognition, machine translation, image
generation, and so on. However, auto-regressive models require a
new output to be added to the end of the current input sequence at
each of multiple time steps. The described techniques, on the other
hand, allow for new outputs to be added at arbitrary positions
within the current input sequence and, in some cases, multiple
outputs to be added at multiple different positions at a single
time step.
[0007] When only a single output is added at each time step, the
additional flexibility afforded to the neural network to select the
position at which the single output is added improves performance
over conventional auto-regressive models without increasing the
number of generation time steps required.
[0008] When multiple outputs can be added at each time step,
because the number of generation time steps for which processing is
required to be performed is reduced, the described techniques allow
for outputs to be generated much quicker (and while using fewer
computational resources) than outputs generated by auto-regressive
models without a significant decrease (and, in some cases, with an
increase) in output generation quality. In other words, by
generating multiple different outputs in parallel at the same time
step, the system can generate the output over fewer generation time
steps and use fewer computational resources than conventional
systems while still generating network outputs that are comparable
in quality to these conventional systems.
[0009] The details of one or more embodiments of the subject matter
described in this specification are set forth in the accompanying
drawings and the description below. Other features, aspects, and
advantages of the subject matter will become apparent from the
description, the drawings, and the claims.
BRIEF DESCRIPTION OF THE DRAWINGS
[0010] FIG. 1 shows an example neural network system.
[0011] FIG. 2 illustrates the generation of an example network
output using the neural network system.
[0012] FIG. 3 is a flow diagram of an example process for
generating a network output.
[0013] FIG. 4 is a flow diagram of an example process for
generating score distributions.
[0014] Like reference numbers and designations in the various
drawings indicate like elements.
DETAILED DESCRIPTION
[0015] This specification describes a system implemented as
computer programs on one or more computers in one or more locations
that generates a network output that includes a respective output
from a vocabulary of outputs at each of multiple locations in an
output order. In some cases, the output is a one-dimensional
sequence, e.g., a sequence of text, while in other cases the output
is a higher-dimensional array, e.g., an image. The vocabulary of
outputs can include each possible output that can be selected when
performing a machine learning task to generate a network
output.
[0016] For example, the system may be a neural machine translation
system. That is, if the network input is a sequence of words in an
original language, e.g., a sentence or phrase, the network output
may be a translation of the network input into a target language,
i.e., a sequence of words in the target language that represents
the sequence of words in the original language.
[0017] As another example, the system may be a speech recognition
system. That is, if the network input is a sequence of audio data
representing a spoken utterance, the network output may be a
sequence of graphemes, characters, or words that represents the
utterance, i.e., is a transcription of the network input.
[0018] As another example, the system may be a natural language
processing system. For example, if the network input is a sequence
of words in an original language, e.g., a sentence or phrase, the
network output may be a summary of the network input in the
original language, i.e., a sequence that has fewer words than the
network input but that retains the essential meaning of the network
input. As another example, if the network input is a sequence of
words that form a question, the network output can be a sequence of
words that form an answer to the question.
[0019] As another example, the system may be part of a
computer-assisted medical diagnosis system. For example, the
network input can be a sequence of data from an electronic medical
record and the network output can be a sequence of predicted
treatments.
[0020] As another example, the system may be part of an image
processing system. For example, the network input can be an image,
i.e., a sequence of color values from the image, and the output can
be a sequence of text that describes the image. As another example,
the network input can describe the context of an image, .e.g., be a
sequence of text, and the network output can be an image that
describes the context.
[0021] FIG. 1 shows an example neural network system 100. The
neural network system 100 is an example of a system implemented as
computer programs on one or more computers in one or more locations
in which the systems, components, and techniques described below
are implemented.
[0022] The system 100 receives a network input 102 processes the
network input 102 to generate a network output 132 for the network
input 102.
[0023] In particular, the system 100 generates the network output
132 over multiple generation time steps.
[0024] At each generation time step, the system 100 selects one or
more new outputs 128 from the vocabulary to be added to the network
output conditioned on the network input 102 and a current partial
network output 122 that has already been generated as of the
generation time step.
[0025] As of any given generation time step, the current partial
network output 122 has zero or more outputs from the vocabulary of
outputs arranged according to a partial output order. In other
words, at the first generation time step, the current partial
network output 122 is empty, i.e., has zero outputs, and, at all
other generation time steps, the current partial network output 122
has the outputs that were previously added at previous generation
time steps.
[0026] In more detail, at each generation time step, the system 100
identifies the current partial network output 122 that has already
been generated as of the generation time step.
[0027] The system 100 then performs one or more insertion
operations to add one or more new outputs 128 to the current
partial network output 122.
[0028] In particular, the system 100 generates, using a decoder
neural network 110 conditioned on (i) at least a portion of the
network input 102 and (ii) the outputs in the current partial
network output 122, a decoder output 112.
[0029] The decoder output 112 defines, for each of a plurality of
insertion locations, a respective score distribution over the
vocabulary of outputs.
[0030] Each insertion location is a different new location in the
partial output order at which there is no output in the current
partial network output, i.e., a new location that can be added to
the partial output order (i) before all of the outputs in the
current partial network output, (ii) between two outputs in the
current partial network output, or (iii) after all of the outputs
in the current partial network output.
[0031] For the first generation time step, there may only be a
single insertion location that adds the first output to the current
partial network output 122, i.e., because for the first generation
time step the current partial network output 122 is empty and there
are no outputs already in the current partial network output
122.
[0032] An output selection engine 120 within the system 100 then
selects, using the decoder output 112, one or more of the insertion
locations and, for each selected insertion location, an inserted
output from the vocabulary, i.e., selects one or more new outputs
128 and one or more insertion locations 128.
[0033] The system 100 then updates the current partial network
output 122 by generating a new partial network output that includes
(i) any outputs in the current partial network output 122 and (ii)
for each selected insertion location, the inserted output from the
vocabulary inserted at the corresponding new location in the
partial output order.
[0034] In some implementations (referred to as "greedy decoding"),
the engine 120 selects only a single insertion location, i.e., adds
a single output to the current output, at each generation time
step. In these implementations, because the engine 120 selects the
insertion location at which the single output will be added rather
than automatically adding the single output to the end of the
current output 122 as would be done by conventional systems, the
quality of the final output can be increased relative to
conventional systems, i.e., because the network output can be
generated according to a more flexible ordering when adding outputs
to the end of the current outputs 122 is insufficient or
suboptimal.
[0035] In some other implementations (referred to as "parallel
decoding"), the engine 120 can select multiple insertion locations,
i.e., add multiple outputs to the current partial output 122, at
any given generation time step. In these implementations, because,
unlike in conventional systems, multiple outputs can be added at a
single generation time step, the number of time steps required to
generate the network output 132 and, therefore, the time and amount
of computational resources required to generate the network output
132, can be greatly reduced.
[0036] Selecting insertion locations and outputs using both greedy
decoding and parallel decoding is described in more detail below
with reference to FIG. 3.
[0037] The system 100 can continue adding outputs at generation
time steps until termination criteria have been reached, i.e.,
until certain criteria involving an end-of-sequence token that is
in the vocabulary are satisfied, as described in more detail below
with reference to FIG. 3.
[0038] That is, at each generation time step, the system 100
checks, using the decoder output 112, to see if the criteria are
satisfied and only adds a new output 128 to the current partial
output 122 if the criteria are not satisfied.
[0039] If the criteria are satisfied, the system 100 determines
that the current partial output 122 is the final output to be
generated, does not add any more outputs to the current partial
output 122, and outputs the current partial output 122 as the final
network output 132.
[0040] In some cases, the decoder neural network 110 is an
attention-based decoder neural network that applies attention over
encoded representations of the network input, i.e., generated by an
encoder neural network (which can be part of the system 100 or an
external system that encodes the network input 102 before the
network input is provided to the system 100), and self-attention
over the outputs in the current partial output 122.
[0041] Such a neural network is described in PCT Application No.
PCT/US2018/034224, filed on May 23, 2018, the entire contents of
which are hereby incorporated by reference herein in their
entirety. Some attention-based decoders apply causal masked
self-attention over the outputs in the current partial output to
prevent outputs at any given position from being influenced by
outputs at future positions. Because the system 100 is able to
insert an output at any insertion location, unlike these other
attention-based decoders, the decoder neural network 110 does not
apply a causal self-attention mask from the decoder so that all
positions can attend to all other positions, as opposed to just
those to the left of the current position. This allows each
decision to condition on the full context of the current partial
output at any given generation time step. Other modifications to
such neural networks that may improve the performance of the neural
network when generating insertion operations are described below
with reference to FIG. 4.
[0042] In other cases however, the decoder neural network 110 is a
different type of neural network that maps network inputs to
network outputs.
[0043] For example, the decoder 110 can be a recurrent neural
network that is conditioned on the current partial output 122
through a recurrent state and that applies attention over encoded
representations of the network input generated by an encoder neural
network.
[0044] As another example, the decoder 110 can be a convolutional
neural network that receives as input the current partial output
and has one or more convolutional neural network layers that are
conditioned on a representation of the network input.
[0045] FIG. 2 shows the generation of an example network output
using the neural network system 100. In the example of FIG. 2, the
network output that is being generated is [three, friends, ate,
lunch, together], where each of the words in the output is selected
from a vocabulary of words.
[0046] Segment 210 of FIG. 2 illustrates the generation of the
network output using "greedy" decoding over six time steps 0-5
shown in in column "t". The current partial output as of any one of
the six time steps is shown in the column labelled "Canvas" while
the insertion operation performed at the time step is shown column
"Insertion." Thus, as can been from segment 210, in greedy decoding
the system adds a single output to the current partial output at
each time step. However, unlike in conventional systems, for
example, at time step 2, the word "friends" is added at location 0,
i.e., to the front of the partial output [ate, together].
[0047] Segment 220 of FIG. 2 illustrates the generation of the
network output using "parallel decoding" over four time steps 0-4.
As can been from segment 220, in parallel decoding the system can
add multiple outputs to the current partial output at a given time
step. For example, at time step 1, "friends" is added at insertion
location 0, to the front of the current partial output, while
"together" is added to insertion location 1. As another example, at
time step 2, "three" is added at insertion location 0, while
"lunch" is added to insertion location 2, i.e., between the words
"ate" and "together."
[0048] FIG. 3 is a flow diagram of an example process 300 for
generating a network output from a network input. For convenience,
the process 300 will be described as being performed by a system of
one or more computers located in one or more locations. For
example, a neural network system, e.g., the neural network system
100 of FIG. 1, appropriately programmed, can perform the process
300.
[0049] The system can perform the process 300 at each of multiple
generation time steps to generate the network output from the
network input. In particular, the system continues performing
process 300 until termination criteria are satisfied at step
306.
[0050] The system identifies the current partial network output
that has already been generated as of the generation time step
(step 302).
[0051] The current partial network output has zero or more outputs
from the vocabulary of outputs arranged according to a partial
output order. In other words, at the first generation time step,
the partial output is empty, i.e., has zero outputs, and, at all
other generation time steps, the partial output has the outputs
that were previously added at previous generation time steps.
[0052] The system generates, using the decoder neural network and
conditioned on (i) at least a portion of the network input and (ii)
any outputs that are in the current partial network output, a
decoder output (step 304).
[0053] The decoder output defines, for each of a plurality of
insertion locations, a respective score distribution over the
vocabulary of outputs. Each insertion location is a different new
location in the partial output order at which there is no output in
the current partial network. For example, at the first time step,
there is only a single new location in the partial output order,
i.e., the location at which the first output can be added to the
network output. At the other generation time steps, the new
locations include (i) the location before any of the outputs in the
current partial network output, (ii) the location after all of the
outputs in the current partial network output, and (iii) if there
is more than one network output in the current partial network
output, the respective locations between each consecutive pair of
network outputs in the current partial network outputs. As a
particular example, if the current partial network output is
[friends, ate, together], the partial network output with the new
locations (denoted by "_") at which a new output could be inserted
shown would be [_, friends, _, ate, _, together, _].
[0054] Generating the decoder output is described in more detail
below with reference to FIG. 4.
[0055] The system then determines, based on the decoder output,
whether the network output should be terminated (step 306) and, if
not, selects, using the decoder output, one or more of the
insertion locations and, for each selected insertion location, an
inserted output from the vocabulary (step 308).
[0056] That is, if system determines that the network output should
be terminated, the system uses the current partial output as the
final network output. If the system determines that the network
output should not be terminated, the system selects one or more
outputs to be added to the current partial output.
[0057] The manner in which the system determines whether the
network output should be terminated and, if not, the manner in
which the system selects the insertion locations depends on whether
the system is configured to perform greedy decoding or parallel
decoding.
[0058] As described above, in greedy decoding the system selects
only a single insertion location, i.e., adds a single output to the
current output, at each generation time step.
[0059] In some of the implementations in which the system performs
greedy decoding, the system can determine to terminate the network
output when the insertion location--output combination with the
highest score across all insertion location--output combinations
includes, as the output in the combination, an end-of-sequence
token. Generally, the end-of-sequence token is a predetermined
token that is added to the vocabulary but will never be produced as
part of a network output and will only be used by the system to
determine when to terminate the network output. The score for a
given insertion location--given output combination refers to the
score for the given output in the score distribution for the given
insertion location as defined by the decoder output.
[0060] In these implementations, when the system determines that
the insertion location--output combination with the highest score
across all insertion location--output combinations does not include
the end-of-sequence token, the system selects only the insertion
location--output combination with the highest score across all
insertion location--output combinations.
[0061] In other words, the system selects, as the insertion
location, the insertion location in the insertion location--output
combination with the highest score across all insertion
location--output combinations and selects, as the inserted output
for the selected insertion location, the output from the insertion
location--output combination with the highest score across all
insertion location--output combinations.
[0062] In others of the implementations in which the system
performs greedy decoding, the system can determine to terminate the
network output only when, for all of the insertion locations, the
output with the highest score is the end-of-sequence token
insertion.
[0063] In other words, the system determines not to terminate the
network output whenever there is at least one insertion location
for which the output with the highest score is not the
end-of-sequence token.
[0064] In these implementations, in response to determining not to
terminate, the system selects only the insertion location--output
combination with the highest score across all insertion
location--output combinations.
[0065] As described above, in parallel decoding the system can
select multiple insertion locations, i.e., add multiple outputs to
the current partial output, at any given generation time step.
[0066] When performing parallel decoding, the system identifies,
from the decoder output and for each insertion location, an output
that has a highest score in the score distribution for the
insertion location. The system then determines to terminate the
network output only when the end-of sequence-token is the output
with the highest score in the score distributions for all of the
insertion locations.
[0067] In other words, the system determines not to terminate when
there is at least one insertion location for which the output with
the highest score is not the end-of-sequence token.
[0068] In response to determining not to terminate, the system
selects each insertion location for which the output with the
highest score is not the end-of-sequence token. For each selected
location, the system then selects the corresponding output that has
the highest score for the selected insertion location. Thus, when
there are multiple insertion locations for which the output that
has the highest score is an output other than the end-of-sequence
token, the system selects multiple insertion locations at the
generation time step.
[0069] The system then generates a new partial network output that
includes (i) any outputs in the current partial network output and
(ii) for each selected insertion location, the inserted output from
the vocabulary inserted at the corresponding new location in the
partial output order (step 310).
[0070] FIG. 4 is a flow diagram of an example process 400 for
generating a decoder output. For convenience, the process 400 will
be described as being performed by a system of one or more
computers located in one or more locations. For example, a neural
network system, e.g., the neural network system 100 of FIG. 1,
appropriately programmed, can perform the process 400.
[0071] The system generates a decoder input for the generation time
step (step 402). As described above, the decoder input conditions
the neural network on the current partial network output as of the
time step and on the network input.
[0072] In particular, as described above, in some implementations
the decoder neural network is a self-attention decoder neural
network or an auto-regressive convolutional decoder neural network
that receives a decoder input that includes the outputs in the
current partial network output arranged according to the partial
output order and encoded representations of the network input.
[0073] In these implementations, the system can, as part of
generating the decoder input, augment the current partial network
output by adding two marker outputs to the current partial network
output in pre-determined positions in the partial output order,
e.g., before the first output in the partial output order and after
the last output in the partial output order. A marker output is a
pre-determined output that is never emitted as part of a network
output and is only used to augment the partial network output.
[0074] The system processes the decoder input using the neural
network generate a respective slot representation for each
insertion location (step 404).
[0075] In particular, at each time step, both self-attention
decoders and auto-regressive convolutional decoders are configured
to generate a respective representation vector for each output in
the current partial output as of the time step.
Because the current partial output has been augmented with the
marker outputs, when the current partial output includes N outputs,
the decoder neural network generates N+2 representations: one for
each of the N outputs and one for each of the marker outputs.
[0076] Moreover, when there are N outputs, there are N+1 possible
insertion locations: (i) one location before any of the outputs in
the current partial network output, (ii) one location after all of
the outputs in the current partial network output, and (iii) if
there is more than one network output in the current partial
network output, the N-1 respective locations between each
consecutive pair of network outputs in the current partial network
outputs.
[0077] To generate the slot representations for the N+1 insertion
locations, for each insertion location, the system generates the
respective slot representation for the insertion location by
concatenating the representation vectors for each adjacent pair of
locations in the partial output order. For example, to generate the
respective slot representation for the insertion location between
the second and third output in the partial output order, the system
would concatenate the representation vectors for the second and
third outputs.
The system generates the score distributions from at least the slot
representations (step 406).
[0078] The system can generate the score distribution from the slot
representations in any of a variety of ways.
[0079] As one example, the system can project a decoder hidden
state matrix generated from the slot representations, i.e., a
matrix that has the slot representations as either the rows or the
columns of the matrix, using a projection matrix to generate a
content-location logit matrix. This projection matrix can be
learned during the training of the decoder neural network.
[0080] The system can then flatten the content-location logit
matrix into a content-location logit vector and apply a softmax
over the content-location logit vector to generate a probability
distribution over all insertion location--output combinations.
[0081] As another example, the system can generate a respective
probability for each insertion location by applying a softmax to
the product of the decoder hidden state matrix and a learned query
vector, i.e., a learned query vector that, when applied to the
decoder hidden state matrix, maps the hidden state matrix to a
vector that includes a respective value for each insertion
location.
[0082] For each insertion location, the system can then project the
slot representation for the location into a score vector that
includes a respective score for each output in the vocabulary using
a projection matrix (that is also learned during the training of
the decoder) and apply a softmax over the score vector to generate
an initial probability for each output in the vocabulary.
[0083] To generate the score distribution for each insertion
location, the system then multiplies each initial probability by
the probability for the location to generate a final probability
for each output in the vocabulary.
[0084] In either of these examples, the system can optionally also
incorporate a bias vector to increase information sharing across
insertion locations. In particular, the system can generate a
context vector by applying max pooling over the slot
representations and then generate the bias vector from the context
vector, where the bias vector is a vector that includes a
respective bias value for each output in the vocabulary.
The system can then generate the decoder output from the bias
vector and the slot representations. For example, prior to
computing a softmax over any set of logits, the system can add the
bias vector to the logits. Generally, incorporating this bias
vector may be useful in providing the decoder neural network with
coverage information, or in propagating count
[0085] information about common outputs that should appear in
multiple places in the network output.
[0086] In order for the system to be effectively used to generate
network outputs, i.e., to generate high-quality decoder outputs
that the system can use to generate high-quality network outputs,
the system trains the neural network on training data to optimize
an objective function.
[0087] One example of an objective function that can be used to
train the neural network is one that uses a soft order reward
framework in order to train the decoder neural network to follow an
"oracle" policy when generating network outputs. In particular, for
any given insertion operation a at a given generation time step
that inserts an output w into an insertion location s in between
position i and position j in the output order, the system can
compute a reward value that is equal to the negative of an order
function mapping actions to real numbers, with lower values
corresponding to better actions.
[0088] The system can then minimize a loss that depends on a KL
divergence between (i) an oracle policy for the generation time
step that is based on the reward values and (ii) the score
distributions generated by the decoder neural network at the given
generation time step.
[0089] In particular, the system can train the neural network to
minimize the following loss using conventional gradient-based
neural network training techniques:
R ( a ) = { - O ( a ) .A-inverted. a .di-elect cons. A * - .infin.
.A-inverted. a A * q oracle ( a ) = exp ( R ( a ) / .tau. ) a '
.di-elect cons. A * exp ( R ( a ' ) / .tau. ) L = KL ( q oracle ||
p ) ##EQU00001##
where A* is the set of valid actions at the given generation time
step, p are the score distributions generated by the neural
network, .tau. is a constant temperature parameter, and O(a) is the
order function. An action is a valid action at a given generation
time step if it results in an output from the actual network output
being inserted into the partial output at one of the insertion
locations available at the generation time step. By using a
different order function, the system can train the neural network
to generate network outputs according to a variety of different
orderings. Some examples of order functions that can be used are
shown in Table 1 below.
TABLE-US-00001 TABLE 1 Order Function O(a) 0 |s - (i + j)/2|
rank(hash(w)) .+-.s .+-.rank(frequency(w)) .+-.rank(length(w))
.+-.rank(w) .+-.log p(a)
[0090] As another example, the system can train the neural network
to minimize a loss function that specifically encourages the neural
network to produce its output in a left-to-right fashion. In this
example, the loss function can be of the form:
loss(x,y)=-log p(y.sub.k+1,k|x,y).
where k is a randomly sampled position in the actual network output
y for the network input x, y is the prefix that includes the first
k outputs of the actual network output, and p is the score assigned
to a given output, insertion location pair by the neural
network.
[0091] As yet another example, the system can train the neural
network for maximal parallelism to encourage a balanced binary tree
ordering in generating the network output. In this example, the
loss function can be of the form:
slot - loss ( x , y ^ , l ) = ? - log p ( y i , l | x , y ^ ) w l (
i ) . loss ( x , y ^ ) = 1 k + 1 i = 0 k slot - loss ( x , y ^ , l
) , ? indicates text missing or illegible when filed
##EQU00002##
where w.sub.l(i) is a weight for insertion location i for a
position l in the network output that is based on the distance of
the insertion location i from the center of the span of outputs
from the target output yet to be produced at position l.
[0092] As yet another example, the system can train the neural
network to assign equal probability mass to each correct action
with no preference for which positions in the network output are
generated first. In this example, the loss function can be of the
form:
slot - loss ( x , y ^ , l ) = 1 ? - ? + 1 ? - log p ( y i , l | x ,
y ^ ) . loss ( x , y ^ ) = 1 k + 1 ? slot - loss ( x , y ^ , l ) ,
? indicates text missing or illegible when filed ##EQU00003##
where i.sub.l is the first output that is yet to be produced at
position l and j.sub.l is the last output yet to be produced at
position l.
[0093] This specification uses the term "configured" in connection
with systems and computer program components. For a system of one
or more computers to be configured to perform particular operations
or actions means that the system has installed on it software,
firmware, hardware, or a combination of them that in operation
cause the system to perform the operations or actions. For one or
more computer programs to be configured to perform particular
operations or actions means that the one or more programs include
instructions that, when executed by data processing apparatus,
cause the apparatus to perform the operations or actions.
[0094] Embodiments of the subject matter and the functional
operations described in this specification can be implemented in
digital electronic circuitry, in tangibly-embodied computer
software or firmware, in computer hardware, including the
structures disclosed in this specification and their structural
equivalents, or in combinations of one or more of them. Embodiments
of the subject matter described in this specification can be
implemented as one or more computer programs, i.e., one or more
modules of computer program instructions encoded on a tangible non
transitory storage medium for execution by, or to control the
operation of, data processing apparatus. The computer storage
medium can be a machine-readable storage device, a machine-readable
storage substrate, a random or serial access memory device, or a
combination of one or more of them. Alternatively or in addition,
the program instructions can be encoded on an artificially
generated propagated signal, e.g., a machine-generated electrical,
optical, or electromagnetic signal, that is generated to encode
information for transmission to suitable receiver apparatus for
execution by a data processing apparatus.
[0095] The term "data processing apparatus" refers to data
processing hardware and encompasses all kinds of apparatus,
devices, and machines for processing data, including by way of
example a programmable processor, a computer, or multiple
processors or computers. The apparatus can also be, or further
include, special purpose logic circuitry, e.g., an FPGA (field
programmable gate array) or an ASIC (application specific
integrated circuit). The apparatus can optionally include, in
addition to hardware, code that creates an execution environment
for computer programs, e.g., code that constitutes processor
firmware, a protocol stack, a database management system, an
operating system, or a combination of one or more of them.
[0096] A computer program, which may also be referred to or
described as a program, software, a software application, an app, a
module, a software module, a script, or code, can be written in any
form of programming language, including compiled or interpreted
languages, or declarative or procedural languages; and it can be
deployed in any form, including as a stand alone program or as a
module, component, subroutine, or other unit suitable for use in a
computing environment. A program may, but need not, correspond to a
file in a file system. A program can be stored in a portion of a
file that holds other programs or data, e.g., one or more scripts
stored in a markup language document, in a single file dedicated to
the program in question, or in multiple coordinated files, e.g.,
files that store one or more modules, sub programs, or portions of
code. A computer program can be deployed to be executed on one
computer or on multiple computers that are located at one site or
distributed across multiple sites and interconnected by a data
communication network.
[0097] In this specification, the term "database" is used broadly
to refer to any collection of data: the data does not need to be
structured in any particular way, or structured at all, and it can
be stored on storage devices in one or more locations. Thus, for
example, the index database can include multiple collections of
data, each of which may be organized and accessed differently.
[0098] Similarly, in this specification the term "engine" is used
broadly to refer to a software-based system, subsystem, or process
that is programmed to perform one or more specific functions.
Generally, an engine will be implemented as one or more software
modules or components, installed on one or more computers in one or
more locations. In some cases, one or more computers will be
dedicated to a particular engine; in other cases, multiple engines
can be installed and running on the same computer or computers.
[0099] The processes and logic flows described in this
specification can be performed by one or more programmable
computers executing one or more computer programs to perform
functions by operating on input data and generating output. The
processes and logic flows can also be performed by special purpose
logic circuitry, e.g., an FPGA or an ASIC, or by a combination of
special purpose logic circuitry and one or more programmed
computers.
[0100] Computers suitable for the execution of a computer program
can be based on general or special purpose microprocessors or both,
or any other kind of central processing unit. Generally, a central
processing unit will receive instructions and data from a read only
memory or a random access memory or both. The elements of a
computer are a central processing unit for performing or executing
instructions and one or more memory devices for storing
instructions and data. The central processing unit and the memory
can be supplemented by, or incorporated in, special purpose logic
circuitry. Generally, a computer will also include, or be
operatively coupled to receive data from or transfer data to, or
both, one or more mass storage devices for storing data, e.g.,
magnetic, magneto optical disks, or optical disks. However, a
computer need not have such devices. Moreover, a computer can be
embedded in another device, e.g., a mobile telephone, a personal
digital assistant (PDA), a mobile audio or video player, a game
console, a Global Positioning System (GPS) receiver, or a portable
storage device, e.g., a universal serial bus (USB) flash drive, to
name just a few.
[0101] Computer readable media suitable for storing computer
program instructions and data include all forms of non volatile
memory, media and memory devices, including by way of example
semiconductor memory devices, e.g., EPROM, EEPROM, and flash memory
devices; magnetic disks, e.g., internal hard disks or removable
disks; magneto optical disks; and CD ROM and DVD-ROM disks.
[0102] To provide for interaction with a user, embodiments of the
subject matter described in this specification can be implemented
on a computer having a display device, e.g., a CRT (cathode ray
tube) or LCD (liquid crystal display) monitor, for displaying
information to the user and a keyboard and a pointing device, e.g.,
a mouse or a trackball, by which the user can provide input to the
computer. Other kinds of devices can be used to provide for
interaction with a user as well; for example, feedback provided to
the user can be any form of sensory feedback, e.g., visual
feedback, auditory feedback, or tactile feedback; and input from
the user can be received in any form, including acoustic, speech,
or tactile input. In addition, a computer can interact with a user
by sending documents to and receiving documents from a device that
is used by the user; for example, by sending web pages to a web
browser on a user's device in response to requests received from
the web browser. Also, a computer can interact with a user by
sending text messages or other forms of message to a personal
device, e.g., a smartphone that is running a messaging application,
and receiving responsive messages from the user in return.
[0103] Data processing apparatus for implementing machine learning
models can also include, for example, special-purpose hardware
accelerator units for processing common and compute-intensive parts
of machine learning training or production, i.e., inference,
workloads.
[0104] Machine learning models can be implemented and deployed
using a machine learning framework, e.g., a TensorFlow framework, a
Microsoft Cognitive Toolkit framework, an Apache Singa framework,
or an Apache MXNet framework.
[0105] Embodiments of the subject matter described in this
specification can be implemented in a computing system that
includes a back end component, e.g., as a data server, or that
includes a middleware component, e.g., an application server, or
that includes a front end component, e.g., a client computer having
a graphical user interface, a web browser, or an app through which
a user can interact with an implementation of the subject matter
described in this specification, or any combination of one or more
such back end, middleware, or front end components. The components
of the system can be interconnected by any form or medium of
digital data communication, e.g., a communication network. Examples
of communication networks include a local area network (LAN) and a
wide area network (WAN), e.g., the Internet.
[0106] The computing system can include clients and servers. A
client and server are generally remote from each other and
typically interact through a communication network. The
relationship of client and server arises by virtue of computer
programs running on the respective computers and having a
client-server relationship to each other. In some embodiments, a
server transmits data, e.g., an HTML page, to a user device, e.g.,
for purposes of displaying data to and receiving user input from a
user interacting with the device, which acts as a client. Data
generated at the user device, e.g., a result of the user
interaction, can be received at the server from the device.
[0107] While this specification contains many specific
implementation details, these should not be construed as
limitations on the scope of any invention or on the scope of what
may be claimed, but rather as descriptions of features that may be
specific to particular embodiments of particular inventions.
Certain features that are described in this specification in the
context of separate embodiments can also be implemented in
combination in a single embodiment. Conversely, various features
that are described in the context of a single embodiment can also
be implemented in multiple embodiments separately or in any
suitable subcombination. Moreover, although features may be
described above as acting in certain combinations and even
initially be claimed as such, one or more features from a claimed
combination can in some cases be excised from the combination, and
the claimed combination may be directed to a subcombination or
variation of a subcombination.
[0108] Similarly, while operations are depicted in the drawings and
recited in the claims in a particular order, this should not be
understood as requiring that such operations be performed in the
particular order shown or in sequential order, or that all
illustrated operations be performed, to achieve desirable results.
In certain circumstances, multitasking and parallel processing may
be advantageous. Moreover, the separation of various system modules
and components in the embodiments described above should not be
understood as requiring such separation in all embodiments, and it
should be understood that the described program components and
systems can generally be integrated together in a single software
product or packaged into multiple software products.
[0109] Particular embodiments of the subject matter have been
described. Other embodiments are within the scope of the following
claims. For example, the actions recited in the claims can be
performed in a different order and still achieve desirable results.
As one example, the processes depicted in the accompanying figures
do not necessarily require the particular order shown, or
sequential order, to achieve desirable results. In some cases,
multitasking and parallel processing may be advantageous.
* * * * *