Visual concept learning from few images
summary
Describes concrete methods: using GAN’s and siamese networks to augment data and verify image pairs, episodic training and meta-learner LSTM that proposes parameter updates, and memory-augmented models that store representations. Reports concrete results: matching networks 98.1% on Omniglot but only ~46–60% on ImageNet mini (one‑/five‑shot), Siamese nets ~92% on Omniglot, and GAN-based hallucination for extra examples. Shows prototypes trained end‑to‑end and evaluation numbers that succeed on synthetic benchmarks but drop on real images.
AI-written summary · about
chapters (15)
- 0:00Few-shot visual concept learning
- 1:10Few-shot learning overview
- 4:50One-shot learning challenges
- 7:19Few-shot learning datasets
- 8:04Data vs metric approaches
- 10:12Prior work in few-shot learning
- 11:16Siamese networks summary
- 14:03Episodes and meta-learning
- 17:17Neural Turing machines
- 18:00Memory networks summary
- 20:07Matching networks summary
- 23:21Meta networks summary
- 26:33Research directions
- 27:38Business applications
- 29:06Question and answer session
