Summary of Loca: Logit Calibration For Knowledge Distillation, by Runming Yang et al.
LoCa: Logit Calibration for Knowledge Distillation
by Runming Yang, Taiqiang Wu, Yujiu Yang
First submitted to arxiv on: 7 Sep 2024
Categories
- Main: Computation and Language (cs.CL)
- 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 Knowledge Distillation (KD), a model compression technique, aims to train better student models by mimicking teacher models. One common approach is to align output logits, but our research reveals a crucial issue: mis-instruction, where student models are misled when predictions based on teacher logits do not match labels. Furthermore, we identify dark knowledge in logits that aids distillation. To address this, we propose Logit Calibration (LoCa), a simple yet effective method that corrects predictions and preserves useful information without requiring additional parameters. LoCa is tested on image classification and text generation tasks, demonstrating significant performance improvements over baselines. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Have you ever wondered how to make artificial intelligence models better? One way is called Knowledge Distillation (KD). It helps smaller student models learn from more experienced teacher models. But we found a problem: the student model might get confused if it’s not following the correct answers. We also discovered that there’s valuable information hidden in the teacher model’s predictions, which can actually help distillation. To fix this, we came up with a new method called Logit Calibration (LoCa). It’s easy to use and doesn’t need any extra resources. We tested LoCa on images and text and found it significantly improved performance. |
Keywords
» Artificial intelligence » Distillation » Image classification » Knowledge distillation » Logits » Model compression » Student model » Teacher model » Text generation
