Large Language Models Struggle to Learn Long-Tail Knowledge
Jul 24, 2023
Speakers
Nikhil Kandpal
Speaker · 0 followers
Haikang Deng
Speaker · 0 followers
Adam Roberts
Speaker · 0 followers
Eric Wallace
Speaker · 2 followers
Colin Raffel
Speaker · 0 followers
About
The internet contains a wealth of knowledge—from the birthdays of historical figures to tutorials on how to code—all of which may be learned by language models. However, there is a huge variability in the number of times a piece of information appears on the web. In this paper, we study the relationship between the knowledge memorized by large language models and the information in their pre-training datasets. In particular, we show that a language model's ability to answer a fact-based question relates to how many documents associated with that question were seen during pre-training. We identify these relevant documents by entity linking pre-training datasets and counting documents that contain the same entities as a given question-answer pair. Our results demonstrate strong correlational and causal relationships between accuracy and relevant document count for numerous question answering datasets (e.g., TriviaQA), pre-training corpora (e.g., ROOTS), and model sizes (e.g., 176B parameters). Moreover, we find that while larger models are better than smaller models at learning long-tail knowledge, we estimate that today's models must be scaled by many orders of magnitude to reach competitive QA performance on questions with little support in the pre-training data. Finally, we show that retrieval-augmentation can reduce the dependence on relevant document count, presenting a promising approach for capturing the long-tail.The internet contains a wealth of knowledge—from the birthdays of historical figures to tutorials on how to code—all of which may be learned by language models. However, there is a huge variability in the number of times a piece of information appears on the web. In this paper, we study the relationship between the knowledge memorized by large language models and the information in their pre-training datasets. In particular, we show that a language model's ability to answer a fact-based question…
Like the format? Trust SlidesLive to capture your next event!
Professional recording and live streaming, delivered globally.
Sharing
Recommended Videos
Presentations on similar topic, category or speaker
Hierarchical Neural Coding for Controllable CAD Model Generation
Watch later
Xiang Xu, …
LeadFL: Client Self-Defense against Model Poisoning in Federated Learning
Watch later
Chaoyi Zhu, …
BiRT: Bio-inspired Replay in Vision Transformers for Continual Learning
Watch later
Efficient Interaction-Aware Interval Analysis of Neural Network Feedback Loops
Watch later
A theory of Continuous Generative Flow Networks
Watch later
Salem Lahlou, …