Language-model pretraining with gradient-disentangled embedding sharing
A method for training a language model comprises (a) receiving vectorized training data as input to a multitask pretraining problem; (b) generating modified vectorized training data based on the vectorized training data, according to an upstream data embedding; (c) emitting pretraining output based on the modified vectorized training data, according to a downstream data embedding equivalent to the upstream data embedding; and (d) adjusting the upstream data embedding and the downstream data embedding by computing, based on the pretraining output, a gradient of the upstream data embedding disentangled from a gradient of the downstream data embedding, thereby advancing the multitask pretraining problem toward a pretrained state.
1. A computer system configured as a language-model training service, the computer system comprising:
a language model hosted in the computer system, the language model including:
an upstream encoder configured to receive vectorized training data and emit modified vectorized training data, the upstream encoder including an upstream data embedding, and
a downstream encoder configured to receive the modified vectorized training data and emit pretraining output, the downstream encoder including a downstream data embedding equivalent to the upstream data embedding; and
pretraining logic configured to adjust the upstream data embedding and the downstream data embedding by computing a gradient of the upstream data embedding disentangled from a gradient of the downstream data embedding, wherein the gradient of the upstream data embedding is computed based on a loss function of the upstream encoder and not on a loss function of the downstream encoder, and wherein the gradient of the downstream data embedding is computed based on the loss function of the upstream encoder and the loss function of the downstream encoder.
2. The training service of claim 1 wherein the vectorized training data comprises non-masked data and the modified vectorized training data comprises masked data, wherein the upstream encoder comprises a generator of the masked data, and wherein the downstream encoder comprises a discriminator operating on the masked data.
3. The training service of claim 2 wherein the pretraining output indicates whether each of a plurality of tokens of the masked data is originally present in the non-masked data or is replaced by the generator.
4. The training service of claim 2 wherein the generator is configured to generate ambiguous corruptions in the non-masked data, and wherein the discriminator is configured to distinguish the ambiguous corruptions from tokens originally present in the non-masked data.
5. The training service of claim 2 wherein each data embedding is a token embedding, and wherein the discriminator is configured to execute replaced token detection (RTD).
6. The training service of claim 2 wherein the generator and the discriminator each comprise a neural network, and wherein the generator has half a depth of the discriminator and a full width of the discriminator.
7. The training service of claim 1 wherein the pretraining logic suppresses back propagation of the gradient of the downstream data embedding into the upstream data embedding.
8. The training service of claim 1 wherein the upstream and downstream encoders are configured to execute collectively a multitask pretraining problem.
9. A language-processing service configured for natural language understanding (NLU), the language-processing service comprising:
a language model including:
an upstream sequence of transformer blocks configured to receive vectorized training data and emit modified vectorized training data during pretraining, the upstream sequence of transformer blocks including an upstream data embedding,
a downstream sequence of transformer blocks configured to receive the modified vectorized training data and emit pretraining output during the pretraining, the downstream sequence of transformer blocks including a downstream data embedding equivalent to the upstream data embedding, wherein pretraining logic operative during the pretraining is configured to adjust the upstream data embedding and the downstream data embedding by computing a gradient of the upstream data embedding disentangled from a gradient of the downstream data embedding, wherein the gradient of the upstream data embedding is computed based on a loss function of the upstream sequence of transformer blocks and not on a loss function of the downstream sequence of transformer blocks, and wherein the gradient of the downstream data embedding is computed based on the loss function of the upstream sequence of transformer blocks and the loss function of the downstream sequence of transformer blocks,
wherein the upstream and downstream sequences of transformer blocks are configured to execute collectively a multitask pretraining problem;
an input module configured to convey language input to the language-processing model; and
an output module configured to expose an output of the language-processing model.
10. The language-processing service of claim 9 wherein the upstream sequence of transformer blocks comprises an upstream encoder, and the downstream sequence of transformer blocks comprises a downstream encoder.
11. The language-processing service of claim 9 wherein the NLU includes one or more of question answering, natural language inference, and named-entity recognition.
12. The language-processing service of claim 9 wherein the at least one of the upstream and downstream sequences of transformer blocks provide disentangled attention over a plurality of encodings.
13. The language-processing service of claim 9 wherein the vectorized training data includes multilingual training data.
14. The language-processing service of claim 9 wherein the vectorized training data comprises non-masked data and the modified vectorized training data comprises masked data, wherein the upstream sequence of transformer blocks comprises a generator of the masked data, and wherein the downstream sequence of transformer blocks comprises a discriminator operating on the masked data.
15. The language-processing service of claim 14 wherein the pretraining output indicates whether each of a plurality of tokens of the masked data is originally present in the non-masked data or is replaced by the generator.
16. A method for training a language model, the method comprising:
receiving vectorized training data in an upstream encoder, as input to a multitask pretraining problem;
generating modified vectorized training data in the upstream encoder based on the vectorized training data, according to an upstream data embedding;
emitting pretraining output from a downstream encoder based on the modified vectorized training data, according to a downstream data embedding equivalent to the upstream data embedding; and
adjusting the upstream data embedding and the downstream data embedding in pretraining logic by computing a gradient of the upstream data embedding disentangled from a gradient of the downstream data embedding, thereby advancing the multitask pretraining problem toward a pretrained state, wherein the gradient of the upstream data embedding is computed based on a loss function of the upstream encoder and not on a loss function of the downstream encoder, and wherein the gradient of the downstream data embedding is computed based on the loss function of the upstream encoder and the loss function of the downstream encoder.
17. The method of claim 16 wherein the vectorized training data comprises non-masked data and the modified vectorized training data comprises masked data.
18. The method of claim 17 wherein the pretraining output indicates whether each of a plurality of tokens of the masked data is originally present in the non-masked data.
19. The method of claim 17 wherein generating the modified vectorized training data includes generating ambiguous corruptions in the non-masked data, and wherein the pretraining output distinguishes the ambiguous corruptions from tokens originally present in the non-masked data.
20. The method of claim 17 wherein each data embedding is a token embedding, and wherein the pretraining output is a product of replaced token detection (RTD).