Retrieval Augmented Models
Retrieval Augmented Models use an external source of data at query / inference time to augment the model.
Some benefits include:
- Reduce the need to encode knowledge in model parameters
- Can be updated on the fly without retraining
- Scale better than large models
- Can be audited and corrected without retraining
- REML-tutorial-slides.pdf
- [2407.12982] Retrieval-Enhanced Machine Learning: Synthesis and Opportunities
Vision
Papers
- Retrieval Augmented Classification for Long-Tail Visual Recognition
- Retrieval-Augmented Transformer for Image Captioning (2022-08-22)
- GitHub - FlagOpen/FlagEmbedding: Dense Retrieval and Retrieval-augmented LLMs
NLP
Papers
- Few-shot Learning with Retrieval Augmented Language Model (2022-08-05)
- Improving language models by retrieving from trillions of tokens (2021)
Ideas
- Knowledge Graph Augmented (GNNs / Transformers to aggregate neighborhood)
- Retrieval Augmented Instance Recognition (Few Shot)