data labeling

_{scroll ↓ to Resources}

Contents

• Tabular data
• Image data
• Text data
• Video data
• Audio data
• Resources

Tabular data

• Use https://www.snorkel.org/ to create labeling heuristics
• semi-supervised learning: when a small part of the dataset is labeled we can generates labels for unlabeled data by leveraging a model trained on a that small set of labeled data using supervised learning. this is called pseudo-labeling.
• unsupervised learning using clustering methods such as K-means
  • Metrics
    • inertia: the sum of the distance of all the points from the centroid. If all the points are close to each other, that means they are similar and it is a good cluster. We aim for a small distance for all points from the centroid.
  • Dunn’s index

Image data

• weak supervision
  • create a lot of noisy labels for each data point based on various heuristics and average them
  • for images the label can be 0 or 1 depending on whether each object of interest is present
• transfer learning
• focal loss

Text data

• zero-shot learning
  • topic generation, summarization
• embedding + clustering (instead of classification head)
• Q: What makes a good custom interface for reviewing LLM outputs? – Hamel’s Blog

Video data

Audio data

Resources

• Разметка данных в Label Studio при помощи GPT-4: интеграция ML Backend / Хабр
• Data Labeling in Machine Learning with Python - Vijaya Kumar Suda

---

Links to this File

table file.tags from [[]] and !outgoing([[]])  AND -"Changelog"