Summary of Random Representations Outperform Online Continually Learned Representations, by Ameya Prabhu et al.
Random Representations Outperform Online Continually Learned Representations
by Ameya Prabhu, Shiven Sinha, Ponnurangam Kumaraguru, Philip H.S. Torr, Ozan Sener, Puneet K. Dokania
First submitted to arxiv on: 13 Feb 2024
Categories
- Main: Computer Vision and Pattern Recognition (cs.CV)
- Secondary: Machine Learning (cs.LG)
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 presents a study on the efficacy of continually learned representations in deep networks. The authors empirically demonstrate that existing online continuously trained networks produce inferior representations compared to simple, predefined random transforms. They propose an approach called RanDumb, which projects raw pixels using a fixed random transform and trains a linear classifier on top without storing exemplars. This method outperforms state-of-the-art continually learned representations across various benchmarks. The study highlights the limitations of representation learning in low-exemplar and online continual learning scenarios, challenging prevailing assumptions about effective representation learning. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary The paper shows that current methods for learning representations don’t work well when new data keeps coming in. Instead, they suggest using a simple trick to improve performance. This “RanDumb” method uses a random transformation on the raw image and then adds a simple classifier on top. It works better than current methods across many different benchmarks. The study also shows that training just a linear classifier on top of pre-trained models is often better than fine-tuning or prompt-tuning. Overall, it challenges our understanding of how to learn good representations in these situations. |
Keywords
* Artificial intelligence * Continual learning * Fine tuning * Prompt * Representation learning
