Summary of An Empirical Analysis Of Forgetting in Pre-trained Models with Incremental Low-rank Updates, by Albin Soutif–cormerais et al.
An Empirical Analysis of Forgetting in Pre-trained Models with Incremental Low-Rank Updates
by Albin Soutif–Cormerais, Simone Magistri, Joost van de Weijer, Andew D. Bagdanov
First submitted to arxiv on: 28 May 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 investigates the impact of Low-Rank Adaptation (LoRA) on the forgetting of pretraining foundation tasks and subsequent downstream tasks. LoRA is a technique used to fine-tune large pretrained neural networks on small target datasets, often using modest hardware. The authors study how varying the rank of LoRA affects forgetting rates for both pretraining and downstream tasks. Their findings suggest that the LoRA rank has a significant impact on forgetting patterns, particularly for vision transformers which exhibit “contextual” forgetting. This is in contrast to residual networks, where this behavior is not observed. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper looks at how we can use big AI models to learn new things without forgetting what we already know. These models are pre-trained using lots of data and then fine-tuned for specific tasks. The researchers want to know if changing the way we fine-tune these models affects how well they remember old things and learn new ones. They found that this method has a big impact on how well the models perform, especially when it comes to vision transformers which seem to forget some of what they learned in a more “context-specific” way. |
Keywords
* Artificial intelligence * Lora * Low rank adaptation * Pretraining
