Difference between revisions of "RETRO"
m |
m |
||
| Line 19: | Line 19: | ||
* [https://www.technologyreview.com/2021/12/08/1041557/deepmind-language-model-beat-others-25-times-size-gpt-3-megatron/ DeepMind says its new language model can beat others 25 times its size | Will Douglas Heaven - MIT Technology Review] ... RETRO (for “Retrieval-Enhanced Transformer”), uses an external memory to look up passages of text on the fly, avoiding some of the costs of training a vast neural network | * [https://www.technologyreview.com/2021/12/08/1041557/deepmind-language-model-beat-others-25-times-size-gpt-3-megatron/ DeepMind says its new language model can beat others 25 times its size | Will Douglas Heaven - MIT Technology Review] ... RETRO (for “Retrieval-Enhanced Transformer”), uses an external memory to look up passages of text on the fly, avoiding some of the costs of training a vast neural network | ||
| − | We enhance auto-regressive language models by conditioning on document chunks retrieved from a large corpus, based on local similarity with preceding tokens. With a $2$ trillion token database, our Retrieval-Enhanced Transformer (Retro) obtains comparable performance to GPT-3 and Jurassic-1 on the Pile, despite using 25× fewer parameters. After fine-tuning, Retro performance translates to downstream knowledge-intensive tasks such as question answering. Retro combines a frozen Bert retriever, a differentiable encoder and a chunked cross-attention mechanism to predict tokens based on an order of magnitude more data than what is typically consumed during training. We typically train Retro from scratch, yet can also rapidly RETROfit pre-trained transformers with retrieval and still achieve good performance. Our work opens up new avenues for improving language models through explicit memory at unprecedented scale. - [https://www.deepmind.com/publications/improving-language-models-by-retrieving-from-trillions-of-tokens RETRO | S. Borgeaud, A. Mensch, J. Hoffmann, & L. Sifre] [[Google | DeepMind]] | + | We enhance auto-regressive language models by conditioning on document chunks retrieved from a large corpus, based on local similarity with preceding tokens. With a $2$ trillion token database, our Retrieval-Enhanced Transformer (Retro) obtains comparable performance to GPT-3 and Jurassic-1 on the Pile, despite using 25× fewer parameters. After [[fine-tuning]], Retro performance translates to downstream knowledge-intensive tasks such as question answering. Retro combines a frozen Bert retriever, a differentiable encoder and a chunked cross-attention mechanism to predict tokens based on an order of magnitude more data than what is typically consumed during training. We typically train Retro from scratch, yet can also rapidly RETROfit pre-trained transformers with retrieval and still achieve good performance. Our work opens up new avenues for improving language models through explicit memory at unprecedented scale. - [https://www.deepmind.com/publications/improving-language-models-by-retrieving-from-trillions-of-tokens RETRO | S. Borgeaud, A. Mensch, J. Hoffmann, & L. Sifre] [[Google | DeepMind]] |
Revision as of 20:24, 16 August 2023
YouTube ... Quora ...Google search ...Google News ...Bing News
- Large Language Model (LLM) ... Natural Language Processing (NLP) ... Generation ... Classification ... Understanding ... Translation ... Tools & Services
- Multimodal Language Models ... GPT-4 ... GPT-5
- Attention Mechanism ... Transformer ... Generative Pre-trained Transformer (GPT) ... GAN ... BERT
- Artificial Intelligence (AI) ... Machine Learning (ML) ... Deep Learning ... Neural Network ... Reinforcement ... Learning Techniques
- Assistants ... Personal Companions ... Agents ... Negotiation ... LangChain
- Generative AI ... Conversational AI ... ChatGPT | OpenAI ... Bing | Microsoft ... Bard | Google ... Claude | Anthropic ... Perplexity ... You ... Ernie | Baidu
- DeepMind says its new language model can beat others 25 times its size | Will Douglas Heaven - MIT Technology Review ... RETRO (for “Retrieval-Enhanced Transformer”), uses an external memory to look up passages of text on the fly, avoiding some of the costs of training a vast neural network
We enhance auto-regressive language models by conditioning on document chunks retrieved from a large corpus, based on local similarity with preceding tokens. With a $2$ trillion token database, our Retrieval-Enhanced Transformer (Retro) obtains comparable performance to GPT-3 and Jurassic-1 on the Pile, despite using 25× fewer parameters. After fine-tuning, Retro performance translates to downstream knowledge-intensive tasks such as question answering. Retro combines a frozen Bert retriever, a differentiable encoder and a chunked cross-attention mechanism to predict tokens based on an order of magnitude more data than what is typically consumed during training. We typically train Retro from scratch, yet can also rapidly RETROfit pre-trained transformers with retrieval and still achieve good performance. Our work opens up new avenues for improving language models through explicit memory at unprecedented scale. - RETRO | S. Borgeaud, A. Mensch, J. Hoffmann, & L. Sifre DeepMind
|
|
