Universal transformers
Methods, systems, and apparatus, including computer programs encoded on computer storage media, for implementing a sequence to sequence model that is recurrent in depth while employing self-attention to combine information from different parts of sequences.
1. A method performed by one or more computers, the method comprising:
receiving an input sequence of elements to be transformed into an output sequence of elements according to a learned transformation, each element of the input sequence having a respective initial input representation;
performing an encoding process including repeatedly revising the input representations in parallel using a same series of encoding operations for each of multiple time steps;
initializing a target sequence of elements each having a respective initial target representation;
performing a decoding process including repeatedly revising the target representations in parallel using two-stage self-attention, wherein a second stage of the two-stage self-attention uses an output generated by the encoding process after repeatedly revising the input representations; and
generating, from a final version of the revised target representations, the output sequence of elements representing the learned transformation of the input sequence of elements.
2. The method of claim 1 , wherein performing the decoding process comprises revising the target representations including conditioning on previous representations generated by the decoding process and on the output generated by the encoding process.
3. The method of claim 1 , wherein encoding process uses only one stage of self-attention.
4. The method of claim 1 , wherein the encoding process uses a transition function comprising a fully-connected neural network layer having a single rectified-linear activation function between two linear transformations.
5. The method of claim 1 , wherein the encoding process uses a transition function that is a separable convolution.
6. The method of claim 1 , wherein the encoding process and the decoding process both use neural networks that are recurrent in depth.
7. The method of claim 1 , wherein the input sequence of elements represents a question and the target sequence of elements represents an answer to the question.
8. The method of claim 1 , wherein the input sequence of elements represents a subject of a sentence and the target sequence represents a predicted verb form for the subject of the sentence.
9. The method of claim 1 , wherein the input sequence of elements represents words in a sentence and the target sequence represents a predicted next word in the sentence.
10. The method of claim 1 , wherein the transformation is a symbol transformation for an algorithmic task.
11. The method of claim 1 , wherein the input sequence represents words in a first language and the target sequence represents a translated version of the input sequence in a second language.
12. A system comprising:
one or more computers and one or more storage devices storing instructions that are operable, when executed by the one or more computers, to cause the one or more computers to perform operations comprising:
receiving an input sequence of elements to be transformed into an output sequence of elements according to a learned transformation, each element of the input sequence having a respective initial input representation;
performing an encoding process including repeatedly revising the input representations in parallel using a same series of encoding operations for each of multiple time steps;
initializing a target sequence of elements each having a respective initial target representation;
performing a decoding process including repeatedly revising the target representations in parallel using two-stage self-attention, wherein a second stage of the two-stage self-attention uses an output generated by the encoding process after repeatedly revising the input representations; and
generating, from a final version of the revised target representations, the output sequence of elements representing the learned transformation of the input sequence of elements.
13. The system of claim 12 , wherein performing the decoding process comprises revising the target representations including conditioning on previous representations generated by the decoding process and on the output generated by the encoding process.
14. The system of claim 12 , wherein encoding process uses only one stage of self-attention.
15. The system of claim 12 , wherein the encoding process uses a transition function comprising a fully-connected neural network layer having a single rectified-linear activation function between two linear transformations.
16. The system of claim 12 , wherein the encoding process uses a transition function that is a separable convolution.
17. The system of claim 12 , wherein the encoding process and the decoding process both use neural networks that are recurrent in depth.
18. The system of claim 12 , wherein the input sequence of elements represents a question and the target sequence of elements represents an answer to the question.
19. The system of claim 12 , wherein the input sequence of elements represents a subject of a sentence and the target sequence represents a predicted verb form for the subject of the sentence.
20. One or more non-transitory computer storage media encoded with computer program instructions that when executed by one or more computers cause the one or more computers to perform operations comprising:
receiving an input sequence of elements to be transformed into an output sequence of elements according to a learned transformation, each element of the input sequence having a respective initial input representation;
performing an encoding process including repeatedly revising the input representations in parallel using a same series of encoding operations for each of multiple time steps;
initializing a target sequence of elements each having a respective initial target representation;
performing a decoding process including repeatedly revising the target representations in parallel using two-stage self-attention, wherein a second stage of the two-stage self-attention uses an output generated by the encoding process after repeatedly revising the input representations; and
generating, from a final version of the revised target representations, the output sequence of elements representing the learned transformation of the input sequence of elements.