Summary of A Fresh Take on Stale Embeddings: Improving Dense Retriever Training with Corrector Networks, by Nicholas Monath et al.
A Fresh Take on Stale Embeddings: Improving Dense Retriever Training with Corrector Networks
by Nicholas Monath, Will Grathwohl, Michael Boratko, Rob Fergus, Andrew McCallum, Manzil Zaheer
First submitted to arxiv on: 3 Sep 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: None
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary This paper addresses challenges in training deep encoders for dense retrieval tasks. With increasing numbers of targets and computationally expensive target encoder models, training becomes even more challenging when combined with the use of cached target embeddings that may be outdated due to ongoing training of target encoder parameters. To overcome these challenges, a simple and highly scalable solution is proposed: training a small parametric corrector network that adjusts stale cached target embeddings to enable an accurate softmax approximation and sampling of up-to-date high-scoring “hard negatives.” The paper theoretically investigates the generalization properties of this approach, relating the complexity of the network, staleness of cached representations, and amount of training data. Experimental results on large benchmark dense retrieval datasets as well as QA with retrieval-augmented language models demonstrate that this approach matches state-of-the-art results even when no target embedding updates are made during training beyond an initial cache from an unsupervised pre-trained model, achieving a 4-80x reduction in re-embedding computational cost. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper helps solve big problems with computers learning to find information. When there’s too much information and the computer is slow to learn, it can get stuck. The solution is simple: add a small “fixer” network that makes old information up-to-date again. This allows the computer to make good guesses about what’s important. The paper looks at how well this fixer works and shows that it does just as well as other approaches, but much faster. |
Keywords
» Artificial intelligence » Embedding » Encoder » Generalization » Softmax » Unsupervised
