Summary of A Second-order Perspective on Model Compositionality and Incremental Learning, by Angelo Porrello and Lorenzo Bonicelli and Pietro Buzzega and Monica Millunzi and Simone Calderara and Rita Cucchiara
A Second-Order Perspective on Model Compositionality and Incremental Learning
by Angelo Porrello, Lorenzo Bonicelli, Pietro Buzzega, Monica Millunzi, Simone Calderara, Rita Cucchiara
First submitted to arxiv on: 25 May 2024
Categories
- Main: Artificial Intelligence (cs.AI)
- 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 fine-tuning of pre-trained models has shown compositional properties, where multiple specialized modules can be combined into a single multi-task model. However, identifying the conditions that promote compositionality remains an open issue. The study proposes a theoretical approach to demystify compositionality in non-linear networks through the second-order Taylor approximation of the loss function. The formulation highlights the importance of staying within the pre-training basin to achieve composable modules. Two incremental training algorithms are proposed: one trains individual models, while the other optimizes the composed model as a whole. The algorithms’ application is probed in incremental classification tasks, highlighting valuable skills such as unlearning and specialization. Code available at https://github.com/aimagelab/mammoth. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Deep learning models can be fine-tuned to work on multiple tasks. A new study tries to understand how these models become better with more training. The researchers looked at a special way of understanding how the model works, called the second-order Taylor approximation. This helped them see that staying close to where the model started is important for making it work well on many tasks. They also came up with two ways to train the model: one trains each part separately, and the other trains the whole model together. These methods were tested in a type of classification problem and showed some useful skills like being able to forget what it learned before. |
Keywords
» Artificial intelligence » Classification » Deep learning » Fine tuning » Loss function » Multi task
