Summary of Semantic Residual Prompts For Continual Learning, by Martin Menabue et al.
Semantic Residual Prompts for Continual Learning
by Martin Menabue, Emanuele Frascaroli, Matteo Boschini, Enver Sangineto, Lorenzo Bonicelli, Angelo Porrello, Simone Calderara
First submitted to arxiv on: 11 Mar 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 The paper introduces a novel approach to continual learning (CL) that addresses the issue of catastrophic forgetting in prompt-tuning methods. The method, which leverages a foundation model called CLIP, selects prompts within a two-level adaptation mechanism to ensure stable class prototypes and adapt a pre-trained Vision Transformer (ViT). The authors propose a residual mechanism to transfer CLIP semantics to ViT layers and demonstrate the effectiveness of their approach on established CL benchmarks, outperforming state-of-the-art CL approaches and zero-shot CLIP tests. The findings hold true even for datasets with domain gaps compared to the pre-training knowledge of the backbone model. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary The paper improves prompt-tuning methods for continual learning by making the selection strategy more stable. It does this by using a foundation model called CLIP, which helps select prompts that work well together. This approach is tested on different types of data and shows better results than other methods. The findings are important because they help us learn from new data without forgetting what we already know. |
Keywords
* Artificial intelligence * Continual learning * Prompt * Semantics * Vision transformer * Vit * Zero shot
