Practical Few-Shot Learning
summary
Presents a practical few‑shot learning recipe that often outperforms complex meta‑learning: fine‑tuning with normalized weight vectors plus cosine scoring. Evaluates matching/prototype and metric methods across benchmarks with 600 runs to show high variance, uses FAISS for scalable ANN search, and demonstrates dynamic class addition in two case studies — medical instrument detection (segmentation + matching) and plastic sorting — where synthetic data fails and few‑shot labeling saves cost.
AI-written summary · about
chapters (11)
- 0:00Few-shot learning
- 1:03Traditional ML problems
- 5:40What is few-shot learning
- 8:09Few-shot approaches
- 12:52Prototype methods
- 16:48Memory and gradient methods
- 19:23Efficiency comparison
- 22:34Medical tools case study
- 25:26Plastic sorting in Africa
- 27:33Resource savings
- 28:00Q&A
related talks
26mFew-Shot Learning in Conversational AI
50mFlexible Neural Networks and the Frontiers of Meta-Learning
24mkeras-fsl: Fast Model Builder for Production-Ready Few...
33mBulk Labelling Techniques
30mMachine Learning Techniques for Class Imbalances & Adversaries
01h 04mFace Verification using Few-shot Deep Learning
38mHow to Test a learning system
23mLearning with Limited Labels & Weak Supervision