Learning with Limited Labels & Weak Supervision
summary
Combines thousands-scale weak labels (clicks, dwell time) with ~10k explicit judgments and shows concrete gains from submodular subset selection and similarity/interleaving heuristics for picking weak examples; introduces per-example trust by scaling gradients and a momentum-damped Adam / differentially weighted optimizer that downweights noisy updates. Evaluated on Spotify homepage user-satisfaction prediction and video question answering, adaptive optimization beats naive mixing and subset selection, improving generalization until a plateau when too much weak data is added.
AI-written summary · about
chapters (11)
- 0:00Intro
- 1:34Data Hungry Neural Models
- 3:11Exploiting Weak Supervision
- 4:38Sources of Weak Supervision
- 7:26Data Mixing II: Similarity Thresholded
- 7:46Data Mixing III: Interleaved Learning
- 8:14Data Mixing IV: Submodular Subset
- 9:35Differentially Weighted Learning
- 11:34Momentum Damped Learning
- 14:16User Satisfaction with Recommendation
- 18:07Experiment II: Impact of Adaptive Training
related talks
33mWeak supervision: a new paradigm for unreliable data
45mActive learning and weak supervision in NLP projects
22mPersonalizing Explainable Recommendations with Multi-objective Contextual Bandits
43mADAM - Bootstrapping a Deep Neural Network Sequence Labeling Model with minimal labelling
37mActive Teaching, Human Learning
30mActive Learning and Human-in-the-Loop Machine Learning
28mPreference Learning from Minimal Human Feedback for Interactive
42mBeyond Adversarial Approaches on Distributional Robustness