Handling the problem of hand-labeled training data with data programming...
summary
Presents using Snorkel-style data programming: write noisy rules and label functions (e.g., emoji signals for sentiment, heuristics for images, medical reports, ECGs), combine them with a generative label model to produce probabilistic labels, then train a discriminative classifier. Shows this pipeline can match or exceed fully hand-labeled training while reducing manual labeling effort by orders of magnitude, including results from Google’s Snorkel Pro experiments.
AI-written summary · about
chapters (12)
- 0:00Hand-labeled data problem
- 3:16Traditional labeling drawbacks
- 5:01Alternative labeling methods
- 11:15Data programming with Snorkel
- 14:50Expert role in data programming
- 20:50Building rule functions
- 25:53Generative vs discriminative models
- 31:48Snorkel tool and workflow
- 33:31Google case study
- 38:51Image and medical analysis
- 43:33Summary
- 43:57Q&A
related talks
33mWeak supervision: a new paradigm for unreliable data
36mProgrammatically creating and managing training data
25mData Labeling as Religious Experience
30mEfficient data labelling with weak supervision
29mAnnotating data the right way
01h 09mFrom Walruses to Code: A Guide to Tackling Unique ML Problems
37mCreating Knowledgebases from unstructured text
19mDoing Machine Learning Without a Data Scientist