Few-Shot Learning in Conversational AI
summary
Shows that humans hit >60% accuracy with just five examples on 3x3 rule tasks while standard ML (decision tree, logistic, MLP) lags and sometimes never learns a simple “five black squares” rule; defines few‑shot as 10 or less examples; uses large pretrained LMs (GPT‑3 175B) plus data augmentation, prompt/pattern methods, meta‑learning, Snorkel‑style rules and realtime active learning with human fallback; flags problems combining few‑shot classes with high‑data classes and inconsistent benchmarking assumptions.
AI-written summary · about
chapters (8)
- 0:00Conversational AI & few-shot learning
- 1:49Motivation for few-shot learning
- 6:51Why few-shot learning matters
- 11:14Few-shot learning algorithms
- 16:15Meta learning with examples
- 19:32Customer care IVAs
- 22:41Challenges & future research
- 24:07Human vs machine NLU
related talks
38mPractical Few-Shot Learning
58mIn-Context Principle Learning from Mistakes
45mZero to Hero: Few Shot Learning + Multi-armed Bandit
26mLLM Inference with Few-Shot Prompting and Chain of Thought
50mFlexible Neural Networks and the Frontiers of Meta-Learning
38mPlaying by the Rules-Based-Systems
41mLearning from multiple modalities, Predicting on unseen tasks
27mFireside Chat